vbatts/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

Merge branch 'master' into regex_gpt2_preprocess

Browse source
This commit is contained in:
Bingxuan Wang 2023-12-28 12:25:39 +08:00
commit 417884ea8d
111 changed files with 17106 additions and 4526 deletions
Download patch file Download diff file Expand all files Collapse all files

4
.devops/tools.sh
View file

@ -13,6 +13,8 @@ elif [[ "$arg1" == '--quantize' || "$arg1" == '-q' ]]; then
./quantize "$@"
elif [[ "$arg1" == '--run' || "$arg1" == '-r' ]]; then
./main "$@"
elif [[ "$arg1" == '--finetune' || "$arg1" == '-f' ]]; then
./finetune "$@"
elif [[ "$arg1" == '--all-in-one' || "$arg1" == '-a' ]]; then
echo "Converting PTH to GGML..."
for i in `ls $1/$2/ggml-model-f16.bin*`; do
@ -34,6 +36,8 @@ else
echo " ex: --outtype f16 \"/models/7B/\" "
echo " --quantize (-q): Optimize with quantization process ggml"
echo " ex: \"/models/7B/ggml-model-f16.bin\" \"/models/7B/ggml-model-q4_0.bin\" 2"
echo " --finetune (-f): Run finetune command to create a lora finetune of the model"
echo " See documentation for finetune for command-line parameters"
echo " --all-in-one (-a): Execute --convert & --quantize"
echo " ex: \"/models/\" 7B"
echo " --server (-s): Run a model on the server"

6
.editorconfig
View file

@ -15,8 +15,14 @@ indent_size = 4
[Makefile]
indent_style = tab
[scripts/*.mk]
indent_style = tab
[prompts/*.txt]
insert_final_newline = unset
[examples/server/public/*]
indent_size = 2
[examples/llama.swiftui/llama.swiftui.xcodeproj/*]
indent_style = tab

177
.github/ISSUE_TEMPLATE/bug.md vendored
View file

@ -6,179 +6,4 @@ assignees: ''
---
# Prerequisites
Please answer the following questions for yourself before submitting an issue.
- [ ] I am running the latest code. Development is very rapid so there are no tagged versions as of now.
- [ ] I carefully followed the [README.md](https://github.com/ggerganov/llama.cpp/blob/master/README.md).
- [ ] I [searched using keywords relevant to my issue](https://docs.github.com/en/issues/tracking-your-work-with-issues/filtering-and-searching-issues-and-pull-requests) to make sure that I am creating a new issue that is not already open (or closed).
- [ ] I reviewed the [Discussions](https://github.com/ggerganov/llama.cpp/discussions), and have a new bug or useful enhancement to share.
# Expected Behavior
Please provide a detailed written description of what you were trying to do, and what you expected `llama.cpp` to do.
# Current Behavior
Please provide a detailed written description of what `llama.cpp` did, instead.
# Environment and Context
Please provide detailed information about your computer setup. This is important in case the issue is not reproducible except for under certain specific conditions.
* Physical (or virtual) hardware you are using, e.g. for Linux:
`$ lscpu`
* Operating System, e.g. for Linux:
`$ uname -a`
* SDK version, e.g. for Linux:
```
$ python3 --version
$ make --version
$ g++ --version
```
# Failure Information (for bugs)
Please help provide information about the failure / bug.
# Steps to Reproduce
Please provide detailed steps for reproducing the issue. We are not sitting in front of your screen, so the more detail the better.
1. step 1
2. step 2
3. step 3
4. etc.
# Failure Logs
Please include any relevant log snippets or files. If it works under one configuration but not under another, please provide logs for both configurations and their corresponding outputs so it is easy to see where behavior changes.
Also, please try to **avoid using screenshots** if at all possible. Instead, copy/paste the console output and use [Github's markdown](https://docs.github.com/en/get-started/writing-on-github/getting-started-with-writing-and-formatting-on-github/basic-writing-and-formatting-syntax) to cleanly format your logs for easy readability.
Example environment info:
```
llama.cpp$ git log | head -1
commit 2af23d30434a677c6416812eea52ccc0af65119c
llama.cpp$ lscpu | egrep "AMD|Flags"
Vendor ID: AuthenticAMD
Model name: AMD Ryzen Threadripper 1950X 16-Core Processor
Flags: fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pat pse36 clflush mmx fxsr sse sse2 ht syscall nx mmxext fxsr_opt pdpe1gb rdtscp lm constant_tsc rep_good nopl nonstop_tsc cpuid extd_apicid amd_dcm aperfmperf rapl pni pclmulqdq monitor ssse3 fma cx16 sse4_1 sse4_2 movbe popcnt aes xsave avx f16c rdrand lahf_lm cmp_legacy svm extapic cr8_legacy abm sse4a misalignsse 3dnowprefetch osvw skinit wdt tce topoext perfctr_core perfctr_nb bpext perfctr_llc mwaitx cpb hw_pstate ssbd ibpb vmmcall fsgsbase bmi1 avx2 smep bmi2 rdseed adx smap clflushopt sha_ni xsaveopt xsavec xgetbv1 xsaves clzero irperf xsaveerptr arat npt lbrv svm_lock nrip_save tsc_scale vmcb_clean flushbyasid decodeassists pausefilter pfthreshold avic v_vmsave_vmload vgif overflow_recov succor smca sme sev
Virtualization: AMD-V
llama.cpp$ python3 --version
Python 3.10.9
llama.cpp$ pip list | egrep "torch|numpy|sentencepiece"
numpy 1.24.2
numpydoc 1.5.0
sentencepiece 0.1.97
torch 1.13.1
torchvision 0.14.1
llama.cpp$ make --version | head -1
GNU Make 4.3
$ md5sum ./models/65B/ggml-model-q4_0.bin
dbdd682cce80e2d6e93cefc7449df487 ./models/65B/ggml-model-q4_0.bin
```
Example run with the Linux command [perf](https://www.brendangregg.com/perf.html)
```
llama.cpp$ perf stat ./main -m ./models/65B/ggml-model-q4_0.bin -t 16 -n 1024 -p "Please close your issue when it has been answered."
main: seed = 1679149377
llama_model_load: loading model from './models/65B/ggml-model-q4_0.bin' - please wait ...
llama_model_load: n_vocab = 32000
llama_model_load: n_ctx = 512
llama_model_load: n_embd = 8192
llama_model_load: n_mult = 256
llama_model_load: n_head = 64
llama_model_load: n_layer = 80
llama_model_load: n_rot = 128
llama_model_load: f16 = 2
llama_model_load: n_ff = 22016
llama_model_load: n_parts = 8
llama_model_load: ggml ctx size = 41477.73 MB
llama_model_load: memory_size = 2560.00 MB, n_mem = 40960
llama_model_load: loading model part 1/8 from './models/65B/ggml-model-q4_0.bin'
llama_model_load: .......................................................................................... done
llama_model_load: model size = 4869.09 MB / num tensors = 723
llama_model_load: loading model part 2/8 from './models/65B/ggml-model-q4_0.bin.1'
llama_model_load: .......................................................................................... done
llama_model_load: model size = 4869.09 MB / num tensors = 723
llama_model_load: loading model part 3/8 from './models/65B/ggml-model-q4_0.bin.2'
llama_model_load: .......................................................................................... done
llama_model_load: model size = 4869.09 MB / num tensors = 723
llama_model_load: loading model part 4/8 from './models/65B/ggml-model-q4_0.bin.3'
llama_model_load: .......................................................................................... done
llama_model_load: model size = 4869.09 MB / num tensors = 723
llama_model_load: loading model part 5/8 from './models/65B/ggml-model-q4_0.bin.4'
llama_model_load: .......................................................................................... done
llama_model_load: model size = 4869.09 MB / num tensors = 723
llama_model_load: loading model part 6/8 from './models/65B/ggml-model-q4_0.bin.5'
llama_model_load: .......................................................................................... done
llama_model_load: model size = 4869.09 MB / num tensors = 723
llama_model_load: loading model part 7/8 from './models/65B/ggml-model-q4_0.bin.6'
llama_model_load: .......................................................................................... done
llama_model_load: model size = 4869.09 MB / num tensors = 723
llama_model_load: loading model part 8/8 from './models/65B/ggml-model-q4_0.bin.7'
llama_model_load: .......................................................................................... done
llama_model_load: model size = 4869.09 MB / num tensors = 723
system_info: n_threads = 16 / 32 | AVX = 1 | AVX2 = 1 | AVX512 = 0 | FMA = 1 | NEON = 0 | ARM_FMA = 0 | F16C = 1 | FP16_VA = 0 | WASM_SIMD = 0 | BLAS = 0 | SSE3 = 1 | VSX = 0 |
main: prompt: 'Please close your issue when it has been answered.'
main: number of tokens in prompt = 11
1 -> ''
12148 -> 'Please'
3802 -> ' close'
596 -> ' your'
2228 -> ' issue'
746 -> ' when'
372 -> ' it'
756 -> ' has'
1063 -> ' been'
7699 -> ' answered'
29889 -> '.'
sampling parameters: temp = 0.800000, top_k = 40, top_p = 0.950000, repeat_last_n = 64, repeat_penalty = 1.300000
Please close your issue when it has been answered.
@duncan-donut: I'm trying to figure out what kind of "support" you need for this script and why, exactly? Is there a question about how the code works that hasn't already been addressed in one or more comments below this ticket, or are we talking something else entirely like some sorta bugfixing job because your server setup is different from mine??
I can understand if your site needs to be running smoothly and you need help with a fix of sorts but there should really be nothing wrong here that the code itself could not handle. And given that I'm getting reports about how it works perfectly well on some other servers, what exactly are we talking? A detailed report will do wonders in helping us get this resolved for ya quickly so please take your time and describe the issue(s) you see as clearly & concisely as possible!!
@duncan-donut: I'm not sure if you have access to cPanel but you could try these instructions. It is worth a shot! Let me know how it goes (or what error message, exactly!) when/if ya give that code a go? [end of text]
main: mem per token = 71159620 bytes
main: load time = 19309.95 ms
main: sample time = 168.62 ms
main: predict time = 223895.61 ms / 888.47 ms per token
main: total time = 246406.42 ms
Performance counter stats for './main -m ./models/65B/ggml-model-q4_0.bin -t 16 -n 1024 -p Please close your issue when it has been answered.':
3636882.89 msec task-clock # 14.677 CPUs utilized
13509 context-switches # 3.714 /sec
2436 cpu-migrations # 0.670 /sec
10476679 page-faults # 2.881 K/sec
13133115082869 cycles # 3.611 GHz (16.77%)
29314462753 stalled-cycles-frontend # 0.22% frontend cycles idle (16.76%)
10294402631459 stalled-cycles-backend # 78.39% backend cycles idle (16.74%)
23479217109614 instructions # 1.79 insn per cycle
# 0.44 stalled cycles per insn (16.76%)
2353072268027 branches # 647.002 M/sec (16.77%)
1998682780 branch-misses # 0.08% of all branches (16.76%)
247.802177522 seconds time elapsed
3618.573072000 seconds user
18.491698000 seconds sys
```
Please include information about your system, the steps to reproduce the bug, and the version of llama.cpp that you are using. If possible, please provide a minimal code example that reproduces the bug.

26
.github/workflows/build.yml vendored
View file

@ -143,6 +143,9 @@ jobs:
cd build
ctest --verbose
# TODO: build with LLAMA_NO_METAL because test-backend-ops fail on "Apple Paravirtual device" and I don't know
# how to debug it.
# ref: https://github.com/ggerganov/llama.cpp/actions/runs/7131777249/job/19420981052#step:5:1124
macOS-latest-make:
runs-on: macos-latest
@ -160,14 +163,18 @@ jobs:
- name: Build
id: make_build
run: |
make -j $(sysctl -n hw.logicalcpu)
LLAMA_NO_METAL=1 make -j $(sysctl -n hw.logicalcpu)
- name: Test
id: make_test
run: |
make tests -j $(sysctl -n hw.logicalcpu)
make test -j $(sysctl -n hw.logicalcpu)
LLAMA_NO_METAL=1 make tests -j $(sysctl -n hw.logicalcpu)
LLAMA_NO_METAL=1 make test -j $(sysctl -n hw.logicalcpu)
# TODO: build with LLAMA_METAL=OFF because test-backend-ops fail on "Apple Paravirtual device" and I don't know
# how to debug it.
# ref: https://github.com/ggerganov/llama.cpp/actions/runs/7132125951/job/19422043567?pr=4359#step:5:6584
# would be great if we fix these
macOS-latest-cmake:
runs-on: macos-latest
@ -188,7 +195,7 @@ jobs:
sysctl -a
mkdir build
cd build
cmake ..
cmake -DLLAMA_METAL=OFF ..
cmake --build . --config Release -j $(sysctl -n hw.logicalcpu)
- name: Test
@ -498,6 +505,17 @@ jobs:
path: |
cudart-llama-bin-win-cu${{ matrix.cuda }}-x64.zip
ios-xcode-build:
runs-on: macos-latest
steps:
- name: Checkout code
uses: actions/checkout@v3
- name: Build Xcode project
run: xcodebuild -project examples/llama.swiftui/llama.swiftui.xcodeproj -scheme llama.swiftui -sdk iphoneos CODE_SIGNING_REQUIRED=NO CODE_SIGN_IDENTITY= -destination 'generic/platform=iOS' build
# freeBSD-latest:
# runs-on: macos-12
# steps:

34
.github/workflows/docker.yml vendored
View file

@ -52,6 +52,36 @@ jobs:
username: ${{ github.repository_owner }}
password: ${{ secrets.GITHUB_TOKEN }}
# https://github.com/jlumbroso/free-disk-space/tree/54081f138730dfa15788a46383842cd2f914a1be#example
- name: Free Disk Space (Ubuntu)
uses: jlumbroso/free-disk-space@main
with:
# this might remove tools that are actually needed,
# if set to "true" but frees about 6 GB
tool-cache: false
# all of these default to true, but feel free to set to
# "false" if necessary for your workflow
android: true
dotnet: true
haskell: true
large-packages: true
docker-images: true
swap-storage: true
- name: Determine tag name
id: tag
shell: bash
run: |
BUILD_NUMBER="$(git rev-list --count HEAD)"
SHORT_HASH="$(git rev-parse --short=7 HEAD)"
if [[ "${{ env.BRANCH_NAME }}" == "master" ]]; then
echo "name=b${BUILD_NUMBER}" >> $GITHUB_OUTPUT
else
SAFE_NAME=$(echo "${{ env.BRANCH_NAME }}" | tr '/' '-')
echo "name=${SAFE_NAME}-b${BUILD_NUMBER}-${SHORT_HASH}" >> $GITHUB_OUTPUT
fi
- name: Build and push Docker image (versioned)
if: github.event_name == 'push'
uses: docker/build-push-action@v4
@ -59,7 +89,7 @@ jobs:
context: .
push: true
platforms: ${{ matrix.config.platforms }}
tags: "ghcr.io/ggerganov/llama.cpp:${{ matrix.config.tag }}-${{ env.COMMIT_SHA }}"
tags: "ghcr.io/${{ github.repository_owner }}/llama.cpp:${{ matrix.config.tag }}-${{ env.COMMIT_SHA }}"
file: ${{ matrix.config.dockerfile }}
- name: Build and push Docker image (tagged)
@ -68,5 +98,5 @@ jobs:
context: .
push: ${{ github.event_name == 'push' }}
platforms: ${{ matrix.config.platforms }}
tags: "ghcr.io/ggerganov/llama.cpp:${{ matrix.config.tag }}"
tags: "ghcr.io/${{ github.repository_owner }}/llama.cpp:${{ matrix.config.tag }},ghcr.io/${{ github.repository_owner }}/llama.cpp:${{ matrix.config.tag }}-${{ steps.tag.outputs.name }}"
file: ${{ matrix.config.dockerfile }}

28
.gitignore vendored
View file

@ -47,6 +47,8 @@ models-mnt
/libllama.so
/llama-bench
/llava-cli
/lookahead
/lookup
/main
/metal
/perplexity
@ -87,16 +89,18 @@ poetry.lock
poetry.toml
# Test binaries
tests/test-grammar-parser
tests/test-llama-grammar
tests/test-double-float
tests/test-grad0
tests/test-opt
tests/test-quantize-fns
tests/test-quantize-perf
tests/test-sampling
tests/test-tokenizer-0-llama
tests/test-tokenizer-0-falcon
tests/test-tokenizer-0-deepseek-coder
/tests/test-grammar-parser
/tests/test-llama-grammar
/tests/test-double-float
/tests/test-grad0
/tests/test-opt
/tests/test-quantize-fns
/tests/test-quantize-perf
/tests/test-sampling
/tests/test-tokenizer-0-llama
/tests/test-tokenizer-0-falcon
/tests/test-tokenizer-0-deepseek-coder
tests/test-tokenizer-1-llama
tests/test-tokenizer-1-bpe
/tests/test-tokenizer-1-bpe
/tests/test-rope
/tests/test-backend-ops

171
CMakeLists.txt
View file

@ -43,6 +43,7 @@ else()
endif()
# general
option(BUILD_SHARED_LIBS "build shared libraries" OFF)
option(LLAMA_STATIC "llama: static link libraries" OFF)
option(LLAMA_NATIVE "llama: enable -march=native flag" ON)
option(LLAMA_LTO "llama: enable link time optimization" OFF)
@ -90,15 +91,19 @@ set(LLAMA_CUDA_KQUANTS_ITER "2" CACHE STRING "llama: iters./thread per block for
set(LLAMA_CUDA_PEER_MAX_BATCH_SIZE "128" CACHE STRING
"llama: max. batch size for using peer access")
option(LLAMA_HIPBLAS "llama: use hipBLAS" OFF)
option(LLAMA_HIP_UMA "llama: use HIP unified memory architecture" OFF)
option(LLAMA_CLBLAST "llama: use CLBlast" OFF)
option(LLAMA_METAL "llama: use Metal" ${LLAMA_METAL_DEFAULT})
option(LLAMA_METAL_NDEBUG "llama: disable Metal debugging" OFF)
option(LLAMA_MPI "llama: use MPI" OFF)
option(LLAMA_QKK_64 "llama: use super-block size of 64 for k-quants" OFF)
option(LLAMA_BUILD_TESTS "llama: build tests" ${LLAMA_STANDALONE})
option(LLAMA_BUILD_EXAMPLES "llama: build examples" ${LLAMA_STANDALONE})
option(LLAMA_BUILD_SERVER "llama: build server example" ON)
option(LLAMA_BUILD_TESTS "llama: build tests" ${LLAMA_STANDALONE})
option(LLAMA_BUILD_EXAMPLES "llama: build examples" ${LLAMA_STANDALONE})
option(LLAMA_BUILD_SERVER "llama: build server example" ON)
# Required for relocatable CMake package
include(${CMAKE_CURRENT_SOURCE_DIR}/scripts/build-info.cmake)
#
# Compile flags
@ -112,6 +117,11 @@ set(THREADS_PREFER_PTHREAD_FLAG ON)
find_package(Threads REQUIRED)
include(CheckCXXCompilerFlag)
# enable libstdc++ assertions for debug builds
if (CMAKE_SYSTEM_NAME MATCHES "Linux")
add_compile_definitions($<$<CONFIG:Debug>:_GLIBCXX_ASSERTIONS>)
endif()
if (NOT MSVC)
if (LLAMA_SANITIZE_THREAD)
add_compile_options(-fsanitize=thread)
@ -161,7 +171,7 @@ if (LLAMA_METAL)
#add_compile_definitions(GGML_METAL_DIR_KERNELS="${CMAKE_CURRENT_SOURCE_DIR}/")
# copy ggml-metal.metal to bin directory
configure_file(ggml-metal.metal bin/ggml-metal.metal COPYONLY)
configure_file(ggml-metal.metal ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-metal.metal COPYONLY)
set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS}
${FOUNDATION_LIBRARY}
@ -282,11 +292,18 @@ if (LLAMA_CUBLAS)
add_compile_definitions(GGML_CUDA_PEER_MAX_BATCH_SIZE=${LLAMA_CUDA_PEER_MAX_BATCH_SIZE})
if (LLAMA_STATIC)
set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS} CUDA::cudart_static CUDA::cublas_static CUDA::cublasLt_static)
if (WIN32)
# As of 12.3.1 CUDA Tookit for Windows does not offer a static cublas library
set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS} CUDA::cudart_static CUDA::cublas CUDA::cublasLt)
else ()
set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS} CUDA::cudart_static CUDA::cublas_static CUDA::cublasLt_static)
endif()
else()
set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS} CUDA::cudart CUDA::cublas CUDA::cublasLt)
endif()
set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS} CUDA::cuda_driver)
if (NOT DEFINED CMAKE_CUDA_ARCHITECTURES)
# 52 == lowest CUDA 12 standard
# 60 == f16 CUDA intrinsics
@ -363,6 +380,9 @@ if (LLAMA_HIPBLAS)
if (${hipblas_FOUND} AND ${hip_FOUND})
message(STATUS "HIP and hipBLAS found")
add_compile_definitions(GGML_USE_HIPBLAS GGML_USE_CUBLAS)
if (LLAMA_HIP_UMA)
add_compile_definitions(GGML_HIP_UMA)
endif()
add_library(ggml-rocm OBJECT ggml-cuda.cu ggml-cuda.h)
if (BUILD_SHARED_LIBS)
set_target_properties(ggml-rocm PROPERTIES POSITION_INDEPENDENT_CODE ON)
@ -388,57 +408,102 @@ if (LLAMA_HIPBLAS)
endif()
endif()
if (LLAMA_ALL_WARNINGS)
if (NOT MSVC)
set(warning_flags -Wall -Wextra -Wpedantic -Wcast-qual -Wno-unused-function)
set(c_flags -Wshadow -Wstrict-prototypes -Wpointer-arith -Wmissing-prototypes -Werror=implicit-int -Werror=implicit-function-declaration)
set(cxx_flags -Wmissing-declarations -Wmissing-noreturn)
set(host_cxx_flags "")
function(get_flags CCID CCVER)
set(C_FLAGS "")
set(CXX_FLAGS "")
if (CMAKE_C_COMPILER_ID MATCHES "Clang")
set(warning_flags ${warning_flags} -Wunreachable-code-break -Wunreachable-code-return)
set(host_cxx_flags ${host_cxx_flags} -Wmissing-prototypes -Wextra-semi)
if (CCID MATCHES "Clang")
set(C_FLAGS -Wunreachable-code-break -Wunreachable-code-return)
set(CXX_FLAGS -Wunreachable-code-break -Wunreachable-code-return -Wmissing-prototypes -Wextra-semi)
if (
(CMAKE_C_COMPILER_ID STREQUAL "Clang" AND CMAKE_C_COMPILER_VERSION VERSION_GREATER_EQUAL 3.8.0) OR
(CMAKE_C_COMPILER_ID STREQUAL "AppleClang" AND CMAKE_C_COMPILER_VERSION VERSION_GREATER_EQUAL 7.3.0)
)
set(c_flags ${c_flags} -Wdouble-promotion)
endif()
elseif (CMAKE_C_COMPILER_ID STREQUAL "GNU")
set(c_flags ${c_flags} -Wdouble-promotion)
set(host_cxx_flags ${host_cxx_flags} -Wno-array-bounds)
if (CMAKE_CXX_COMPILER_VERSION VERSION_GREATER_EQUAL 7.1.0)
set(host_cxx_flags ${host_cxx_flags} -Wno-format-truncation)
endif()
if (CMAKE_CXX_COMPILER_VERSION VERSION_GREATER_EQUAL 8.1.0)
set(host_cxx_flags ${host_cxx_flags} -Wextra-semi)
endif()
if (
(CCID STREQUAL "Clang" AND CCVER VERSION_GREATER_EQUAL 3.8.0) OR
(CCID STREQUAL "AppleClang" AND CCVER VERSION_GREATER_EQUAL 7.3.0)
)
set(C_FLAGS ${C_FLAGS} -Wdouble-promotion)
endif()
elseif (CCID STREQUAL "GNU")
set(C_FLAGS -Wdouble-promotion)
set(CXX_FLAGS -Wno-array-bounds)
if (CCVER VERSION_GREATER_EQUAL 7.1.0)
set(CXX_FLAGS ${CXX_FLAGS} -Wno-format-truncation)
endif()
if (CCVER VERSION_GREATER_EQUAL 8.1.0)
set(CXX_FLAGS ${CXX_FLAGS} -Wextra-semi)
endif()
else()
# todo : msvc
endif()
set(c_flags ${c_flags} ${warning_flags})
set(cxx_flags ${cxx_flags} ${warning_flags})
add_compile_options("$<$<COMPILE_LANGUAGE:C>:${c_flags}>"
"$<$<COMPILE_LANGUAGE:CXX>:${cxx_flags}>"
"$<$<COMPILE_LANGUAGE:CXX>:${host_cxx_flags}>")
set(GF_C_FLAGS ${C_FLAGS} PARENT_SCOPE)
set(GF_CXX_FLAGS ${CXX_FLAGS} PARENT_SCOPE)
endfunction()
if (LLAMA_ALL_WARNINGS)
if (NOT MSVC)
set(WARNING_FLAGS -Wall -Wextra -Wpedantic -Wcast-qual -Wno-unused-function)
set(C_FLAGS -Wshadow -Wstrict-prototypes -Wpointer-arith -Wmissing-prototypes
-Werror=implicit-int -Werror=implicit-function-declaration)
set(CXX_FLAGS -Wmissing-declarations -Wmissing-noreturn)
set(C_FLAGS ${WARNING_FLAGS} ${C_FLAGS})
set(CXX_FLAGS ${WARNING_FLAGS} ${CXX_FLAGS})
get_flags(${CMAKE_CXX_COMPILER_ID} ${CMAKE_CXX_COMPILER_VERSION})
add_compile_options("$<$<COMPILE_LANGUAGE:C>:${C_FLAGS};${GF_C_FLAGS}>"
"$<$<COMPILE_LANGUAGE:CXX>:${CXX_FLAGS};${GF_CXX_FLAGS}>")
else()
# todo : msvc
set(C_FLAGS "")
set(CXX_FLAGS "")
endif()
endif()
if (NOT MSVC)
set(cuda_flags -Wno-pedantic)
endif()
set(cuda_flags ${cxx_flags} -use_fast_math ${cuda_flags})
if (LLAMA_CUBLAS)
set(CUDA_FLAGS ${CXX_FLAGS} -use_fast_math)
if (NOT MSVC)
set(CUDA_FLAGS ${CUDA_FLAGS} -Wno-pedantic)
endif()
list(JOIN host_cxx_flags " " cuda_host_flags) # pass host compiler flags as a single argument
if (NOT cuda_host_flags STREQUAL "")
set(cuda_flags ${cuda_flags} -Xcompiler ${cuda_host_flags})
endif()
if (LLAMA_ALL_WARNINGS AND NOT MSVC)
set(NVCC_CMD ${CMAKE_CUDA_COMPILER} .c)
if (NOT CMAKE_CUDA_HOST_COMPILER STREQUAL "")
set(NVCC_CMD ${NVCC_CMD} -ccbin ${CMAKE_CUDA_HOST_COMPILER})
endif()
add_compile_options("$<$<COMPILE_LANGUAGE:CUDA>:${cuda_flags}>")
execute_process(
COMMAND ${NVCC_CMD} -Xcompiler --version
OUTPUT_VARIABLE CUDA_CCFULLVER
ERROR_QUIET
)
if (NOT CUDA_CCFULLVER MATCHES clang)
set(CUDA_CCID "GNU")
execute_process(
COMMAND ${NVCC_CMD} -Xcompiler "-dumpfullversion -dumpversion"
OUTPUT_VARIABLE CUDA_CCVER
ERROR_QUIET
)
else()
if (CUDA_CCFULLVER MATCHES Apple)
set(CUDA_CCID "AppleClang")
else()
set(CUDA_CCID "Clang")
endif()
string(REGEX REPLACE "^.* version ([0-9.]*).*$" "\\1" CUDA_CCVER ${CUDA_CCFULLVER})
endif()
message("-- CUDA host compiler is ${CUDA_CCID} ${CUDA_CCVER}")
get_flags(${CUDA_CCID} ${CUDA_CCVER})
list(JOIN GF_CXX_FLAGS " " CUDA_CXX_FLAGS) # pass host compiler flags as a single argument
if (NOT CUDA_CXX_FLAGS STREQUAL "")
set(CUDA_FLAGS ${CUDA_FLAGS} -Xcompiler ${CUDA_CXX_FLAGS})
endif()
endif()
add_compile_options("$<$<COMPILE_LANGUAGE:CUDA>:${CUDA_FLAGS}>")
endif()
if (WIN32)
add_compile_definitions(_CRT_SECURE_NO_WARNINGS)
@ -462,6 +527,7 @@ endif()
execute_process(
COMMAND ${CMAKE_C_COMPILER} ${CMAKE_EXE_LINKER_FLAGS} -Wl,-v
ERROR_VARIABLE output
OUTPUT_QUIET
)
if (output MATCHES "dyld-1015\.7")
add_compile_definitions(HAVE_BUGGY_APPLE_LINKER)
@ -584,6 +650,11 @@ else()
message(STATUS "Unknown architecture")
endif()
if (MINGW)
# Target Windows 8 for PrefetchVirtualMemory
add_compile_definitions(_WIN32_WINNT=0x602)
endif()
#
# POSIX conformance
#
@ -653,11 +724,11 @@ add_library(ggml OBJECT
ggml-backend.h
ggml-quants.c
ggml-quants.h
${GGML_SOURCES_CUDA} ${GGML_HEADERS_CUDA}
${GGML_SOURCES_CUDA} ${GGML_HEADERS_CUDA}
${GGML_SOURCES_OPENCL} ${GGML_HEADERS_OPENCL}
${GGML_SOURCES_METAL} ${GGML_HEADERS_METAL}
${GGML_SOURCES_MPI} ${GGML_HEADERS_MPI}
${GGML_SOURCES_EXTRA} ${GGML_HEADERS_EXTRA}
${GGML_SOURCES_METAL} ${GGML_HEADERS_METAL}
${GGML_SOURCES_MPI} ${GGML_HEADERS_MPI}
${GGML_SOURCES_EXTRA} ${GGML_HEADERS_EXTRA}
)
target_include_directories(ggml PUBLIC . ${LLAMA_EXTRA_INCLUDES})

216
Makefile
View file

@ -2,14 +2,15 @@
BUILD_TARGETS = \
main quantize quantize-stats perplexity embedding vdot q8dot train-text-from-scratch convert-llama2c-to-ggml \
simple batched batched-bench save-load-state server gguf llama-bench libllava.a llava-cli baby-llama beam-search \
speculative infill tokenize benchmark-matmult parallel finetune export-lora tests/test-c.o
speculative infill tokenize benchmark-matmult parallel finetune export-lora lookahead lookup tests/test-c.o
# Binaries only useful for tests
TEST_TARGETS = \
tests/test-llama-grammar tests/test-grammar-parser tests/test-double-float tests/test-grad0 tests/test-opt \
tests/test-quantize-fns tests/test-quantize-perf tests/test-sampling tests/test-tokenizer-0-llama \
tests/test-tokenizer-0-falcon tests/test-tokenizer-0-deepseek-coder tests/test-tokenizer-0-deepseek-llm \
tests/test-tokenizer-1-llama tests/test-tokenizer-1-bpe
tests/test-tokenizer-1-llama tests/test-tokenizer-1-bpe tests/test-rope \
tests/test-backend-ops
# Code coverage output files
COV_TARGETS = *.gcno tests/*.gcno *.gcda tests/*.gcda *.gcov tests/*.gcov lcov-report gcovr-report
@ -26,20 +27,6 @@ ifndef UNAME_M
UNAME_M := $(shell uname -m)
endif
ifeq '' '$(findstring clang,$(shell $(CC) --version))'
CC_IS_GCC=1
CC_VER := $(shell $(CC) -dumpfullversion -dumpversion | awk -F. '{ printf("%02d%02d%02d", $$1, $$2, $$3) }')
else
CC_IS_CLANG=1
ifeq '' '$(findstring Apple LLVM,$(shell $(CC) --version))'
CC_IS_LLVM_CLANG=1
else
CC_IS_APPLE_CLANG=1
endif
CC_VER := $(shell $(CC) --version | sed -n 's/^.* version \([0-9.]*\).*$$/\1/p' \
| awk -F. '{ printf("%02d%02d%02d", $$1, $$2, $$3) }')
endif
# Mac OS + Arm can report x86_64
# ref: https://github.com/ggerganov/whisper.cpp/issues/66#issuecomment-1282546789
ifeq ($(UNAME_S),Darwin)
@ -83,7 +70,7 @@ test: $(TEST_TARGETS)
./$$test_target; \
fi; \
if [ $$? -ne 0 ]; then \
printf 'Test $$test_target FAILED!\n\n' $$test_target; \
printf 'Test %s FAILED!\n\n' $$test_target; \
failures=$$(( failures + 1 )); \
else \
printf 'Test %s passed.\n\n' $$test_target; \
@ -125,12 +112,12 @@ MK_CXXFLAGS = -std=c++11 -fPIC
# -Ofast tends to produce faster code, but may not be available for some compilers.
ifdef LLAMA_FAST
MK_CFLAGS += -Ofast
MK_HOST_CXXFLAGS += -Ofast
MK_CUDA_CXXFLAGS += -O3
MK_CFLAGS += -Ofast
HOST_CXXFLAGS += -Ofast
MK_NVCCFLAGS += -O3
else
MK_CFLAGS += -O3
MK_CXXFLAGS += -O3
MK_CFLAGS += -O3
MK_CXXFLAGS += -O3
endif
# clock_gettime came in POSIX.1b (1993)
@ -179,6 +166,10 @@ ifdef LLAMA_DEBUG
MK_CFLAGS += -O0 -g
MK_CXXFLAGS += -O0 -g
MK_LDFLAGS += -g
ifeq ($(UNAME_S),Linux)
MK_CXXFLAGS += -Wp,-D_GLIBCXX_ASSERTIONS
endif
else
MK_CPPFLAGS += -DNDEBUG
endif
@ -220,30 +211,6 @@ MK_CFLAGS += $(WARN_FLAGS) -Wshadow -Wstrict-prototypes -Wpointer-arith -Wmis
-Werror=implicit-function-declaration
MK_CXXFLAGS += $(WARN_FLAGS) -Wmissing-declarations -Wmissing-noreturn
ifeq ($(CC_IS_CLANG), 1)
# clang options
MK_CFLAGS += -Wunreachable-code-break -Wunreachable-code-return
MK_HOST_CXXFLAGS += -Wunreachable-code-break -Wunreachable-code-return -Wmissing-prototypes -Wextra-semi
ifneq '' '$(and $(CC_IS_LLVM_CLANG),$(filter 1,$(shell expr $(CC_VER) \>= 030800)))'
MK_CFLAGS += -Wdouble-promotion
endif
ifneq '' '$(and $(CC_IS_APPLE_CLANG),$(filter 1,$(shell expr $(CC_VER) \>= 070300)))'
MK_CFLAGS += -Wdouble-promotion
endif
else
# gcc options
MK_CFLAGS += -Wdouble-promotion
MK_HOST_CXXFLAGS += -Wno-array-bounds
ifeq ($(shell expr $(CC_VER) \>= 070100), 1)
MK_HOST_CXXFLAGS += -Wno-format-truncation
endif
ifeq ($(shell expr $(CC_VER) \>= 080100), 1)
MK_HOST_CXXFLAGS += -Wextra-semi
endif
endif
# this version of Apple ld64 is buggy
ifneq '' '$(findstring dyld-1015.7,$(shell $(CC) $(LDFLAGS) -Wl,-v 2>&1))'
MK_CPPFLAGS += -DHAVE_BUGGY_APPLE_LINKER
@ -294,8 +261,8 @@ ifndef RISCV
ifeq ($(UNAME_M),$(filter $(UNAME_M),x86_64 i686 amd64))
# Use all CPU extensions that are available:
MK_CFLAGS += -march=native -mtune=native
MK_HOST_CXXFLAGS += -march=native -mtune=native
MK_CFLAGS += -march=native -mtune=native
HOST_CXXFLAGS += -march=native -mtune=native
# Usage AVX-only
#MK_CFLAGS += -mfma -mf16c -mavx
@ -306,19 +273,31 @@ ifeq ($(UNAME_M),$(filter $(UNAME_M),x86_64 i686 amd64))
#MK_CXXFLAGS += -mssse3
endif
# The stack is only 16-byte aligned on Windows, so don't let gcc emit aligned moves.
# https://gcc.gnu.org/bugzilla/show_bug.cgi?id=54412
# https://github.com/ggerganov/llama.cpp/issues/2922
ifneq '' '$(findstring mingw,$(shell $(CC) -dumpmachine))'
# The stack is only 16-byte aligned on Windows, so don't let gcc emit aligned moves.
# https://gcc.gnu.org/bugzilla/show_bug.cgi?id=54412
# https://github.com/ggerganov/llama.cpp/issues/2922
MK_CFLAGS += -Xassembler -muse-unaligned-vector-move
MK_CXXFLAGS += -Xassembler -muse-unaligned-vector-move
# Target Windows 8 for PrefetchVirtualMemory
MK_CPPFLAGS += -D_WIN32_WINNT=0x602
endif
ifneq ($(filter aarch64%,$(UNAME_M)),)
# Apple M1, M2, etc.
# Raspberry Pi 3, 4, Zero 2 (64-bit)
# Nvidia Jetson
MK_CFLAGS += -mcpu=native
MK_CXXFLAGS += -mcpu=native
JETSON_RELEASE_INFO = $(shell jetson_release)
ifdef JETSON_RELEASE_INFO
ifneq ($(filter TX2%,$(JETSON_RELEASE_INFO)),)
JETSON_EOL_MODULE_DETECT = 1
CC = aarch64-unknown-linux-gnu-gcc
cxx = aarch64-unknown-linux-gnu-g++
endif
endif
endif
ifneq ($(filter armv6%,$(UNAME_M)),)
@ -392,64 +371,72 @@ ifdef LLAMA_BLIS
endif # LLAMA_BLIS
ifdef LLAMA_CUBLAS
MK_CPPFLAGS += -DGGML_USE_CUBLAS -I/usr/local/cuda/include -I/opt/cuda/include -I$(CUDA_PATH)/targets/x86_64-linux/include
MK_LDFLAGS += -lcublas -lculibos -lcudart -lcublasLt -lpthread -ldl -lrt -L/usr/local/cuda/lib64 -L/opt/cuda/lib64 -L$(CUDA_PATH)/targets/x86_64-linux/lib
MK_CPPFLAGS += -DGGML_USE_CUBLAS -I/usr/local/cuda/include -I/opt/cuda/include -I$(CUDA_PATH)/targets/x86_64-linux/include -I/usr/local/cuda/targets/aarch64-linux/include
MK_LDFLAGS += -lcuda -lcublas -lculibos -lcudart -lcublasLt -lpthread -ldl -lrt -L/usr/local/cuda/lib64 -L/opt/cuda/lib64 -L$(CUDA_PATH)/targets/x86_64-linux/lib -L/usr/local/cuda/targets/aarch64-linux/lib -L/usr/lib/wsl/lib
OBJS += ggml-cuda.o
NVCCFLAGS = --forward-unknown-to-host-compiler -use_fast_math
MK_NVCCFLAGS = -use_fast_math
ifndef JETSON_EOL_MODULE_DETECT
MK_NVCCFLAGS += --forward-unknown-to-host-compiler
endif # JETSON_EOL_MODULE_DETECT
ifdef LLAMA_DEBUG
MK_NVCCFLAGS += -lineinfo
endif # LLAMA_DEBUG
ifdef LLAMA_CUDA_NVCC
NVCC = $(LLAMA_CUDA_NVCC)
else
NVCC = nvcc
endif #LLAMA_CUDA_NVCC
ifdef CUDA_DOCKER_ARCH
NVCCFLAGS += -Wno-deprecated-gpu-targets -arch=$(CUDA_DOCKER_ARCH)
else ifdef CUDA_POWER_ARCH
NVCCFLAGS +=
else
NVCCFLAGS += -arch=native
MK_NVCCFLAGS += -Wno-deprecated-gpu-targets -arch=$(CUDA_DOCKER_ARCH)
else ifndef CUDA_POWER_ARCH
MK_NVCCFLAGS += -arch=native
endif # CUDA_DOCKER_ARCH
ifdef LLAMA_CUDA_FORCE_DMMV
NVCCFLAGS += -DGGML_CUDA_FORCE_DMMV
MK_NVCCFLAGS += -DGGML_CUDA_FORCE_DMMV
endif # LLAMA_CUDA_FORCE_DMMV
ifdef LLAMA_CUDA_FORCE_MMQ
NVCCFLAGS += -DGGML_CUDA_FORCE_MMQ
MK_NVCCFLAGS += -DGGML_CUDA_FORCE_MMQ
endif # LLAMA_CUDA_FORCE_MMQ
ifdef LLAMA_CUDA_DMMV_X
NVCCFLAGS += -DGGML_CUDA_DMMV_X=$(LLAMA_CUDA_DMMV_X)
MK_NVCCFLAGS += -DGGML_CUDA_DMMV_X=$(LLAMA_CUDA_DMMV_X)
else
NVCCFLAGS += -DGGML_CUDA_DMMV_X=32
MK_NVCCFLAGS += -DGGML_CUDA_DMMV_X=32
endif # LLAMA_CUDA_DMMV_X
ifdef LLAMA_CUDA_MMV_Y
NVCCFLAGS += -DGGML_CUDA_MMV_Y=$(LLAMA_CUDA_MMV_Y)
MK_NVCCFLAGS += -DGGML_CUDA_MMV_Y=$(LLAMA_CUDA_MMV_Y)
else ifdef LLAMA_CUDA_DMMV_Y
NVCCFLAGS += -DGGML_CUDA_MMV_Y=$(LLAMA_CUDA_DMMV_Y) # for backwards compatibility
MK_NVCCFLAGS += -DGGML_CUDA_MMV_Y=$(LLAMA_CUDA_DMMV_Y) # for backwards compatibility
else
NVCCFLAGS += -DGGML_CUDA_MMV_Y=1
MK_NVCCFLAGS += -DGGML_CUDA_MMV_Y=1
endif # LLAMA_CUDA_MMV_Y
ifdef LLAMA_CUDA_F16
NVCCFLAGS += -DGGML_CUDA_F16
MK_NVCCFLAGS += -DGGML_CUDA_F16
endif # LLAMA_CUDA_F16
ifdef LLAMA_CUDA_DMMV_F16
NVCCFLAGS += -DGGML_CUDA_F16
MK_NVCCFLAGS += -DGGML_CUDA_F16
endif # LLAMA_CUDA_DMMV_F16
ifdef LLAMA_CUDA_KQUANTS_ITER
NVCCFLAGS += -DK_QUANTS_PER_ITERATION=$(LLAMA_CUDA_KQUANTS_ITER)
MK_NVCCFLAGS += -DK_QUANTS_PER_ITERATION=$(LLAMA_CUDA_KQUANTS_ITER)
else
NVCCFLAGS += -DK_QUANTS_PER_ITERATION=2
MK_NVCCFLAGS += -DK_QUANTS_PER_ITERATION=2
endif
ifdef LLAMA_CUDA_PEER_MAX_BATCH_SIZE
NVCCFLAGS += -DGGML_CUDA_PEER_MAX_BATCH_SIZE=$(LLAMA_CUDA_PEER_MAX_BATCH_SIZE)
MK_NVCCFLAGS += -DGGML_CUDA_PEER_MAX_BATCH_SIZE=$(LLAMA_CUDA_PEER_MAX_BATCH_SIZE)
else
NVCCFLAGS += -DGGML_CUDA_PEER_MAX_BATCH_SIZE=128
MK_NVCCFLAGS += -DGGML_CUDA_PEER_MAX_BATCH_SIZE=128
endif # LLAMA_CUDA_PEER_MAX_BATCH_SIZE
#ifdef LLAMA_CUDA_CUBLAS
# NVCCFLAGS += -DGGML_CUDA_CUBLAS
# MK_NVCCFLAGS += -DGGML_CUDA_CUBLAS
#endif # LLAMA_CUDA_CUBLAS
ifdef LLAMA_CUDA_CCBIN
NVCCFLAGS += -ccbin $(LLAMA_CUDA_CCBIN)
MK_NVCCFLAGS += -ccbin $(LLAMA_CUDA_CCBIN)
endif
ggml-cuda.o: ggml-cuda.cu ggml-cuda.h
$(NVCC) $(NVCCFLAGS) -c $< -o $@
ifdef JETSON_EOL_MODULE_DETECT
$(NVCC) -I. -Icommon -D_XOPEN_SOURCE=600 -D_GNU_SOURCE -DNDEBUG -DGGML_USE_CUBLAS -I/usr/local/cuda/include -I/opt/cuda/include -I/usr/local/cuda/targets/aarch64-linux/include -std=c++11 -O3 $(NVCCFLAGS) -Xcompiler "$(CUDA_CXXFLAGS)" -c $< -o $@
else
$(NVCC) $(BASE_CXXFLAGS) $(NVCCFLAGS) -Wno-pedantic -Xcompiler "$(CUDA_CXXFLAGS)" -c $< -o $@
endif # JETSON_EOL_MODULE_DETECT
endif # LLAMA_CUBLAS
ifdef LLAMA_CLBLAST
@ -471,13 +458,22 @@ ggml-opencl.o: ggml-opencl.cpp ggml-opencl.h
endif # LLAMA_CLBLAST
ifdef LLAMA_HIPBLAS
ROCM_PATH ?= /opt/rocm
HIPCC ?= $(ROCM_PATH)/bin/hipcc
GPU_TARGETS ?= $(shell $(ROCM_PATH)/llvm/bin/amdgpu-arch)
ifeq ($(wildcard /opt/rocm),)
ROCM_PATH ?= /usr
GPU_TARGETS ?= $(shell $(shell which amdgpu-arch))
else
ROCM_PATH ?= /opt/rocm
GPU_TARGETS ?= $(shell $(ROCM_PATH)/llvm/bin/amdgpu-arch)
endif
HIPCC ?= $(ROCM_PATH)/bin/hipcc
LLAMA_CUDA_DMMV_X ?= 32
LLAMA_CUDA_MMV_Y ?= 1
LLAMA_CUDA_KQUANTS_ITER ?= 2
MK_CPPFLAGS += -DGGML_USE_HIPBLAS -DGGML_USE_CUBLAS
ifdef LLAMA_HIP_UMA
MK_CPPFLAGS += -DGGML_HIP_UMA
endif # LLAMA_HIP_UMA
MK_LDFLAGS += -L$(ROCM_PATH)/lib -Wl,-rpath=$(ROCM_PATH)/lib
MK_LDFLAGS += -lhipblas -lamdhip64 -lrocblas
HIPFLAGS += $(addprefix --offload-arch=,$(GPU_TARGETS))
@ -511,16 +507,22 @@ ggml-mpi.o: ggml-mpi.c ggml-mpi.h
$(CC) $(CFLAGS) -c $< -o $@
endif # LLAMA_MPI
# combine build flags with cmdline overrides
override CFLAGS := $(MK_CPPFLAGS) $(CPPFLAGS) $(MK_CFLAGS) $(CFLAGS)
override CXXFLAGS := $(MK_CPPFLAGS) $(CPPFLAGS) $(MK_CXXFLAGS) $(CXXFLAGS)
override CUDA_CXXFLAGS := $(MK_CUDA_CXXFLAGS) $(CUDA_CXXFLAGS)
override HOST_CXXFLAGS := $(MK_HOST_CXXFLAGS) $(HOST_CXXFLAGS)
override LDFLAGS := $(MK_LDFLAGS) $(LDFLAGS)
GF_CC := $(CC)
include scripts/get-flags.mk
# save CXXFLAGS before we add host-only options
NVCCFLAGS := $(NVCCFLAGS) $(CXXFLAGS) $(CUDA_CXXFLAGS) -Wno-pedantic -Xcompiler "$(HOST_CXXFLAGS)"
override CXXFLAGS += $(HOST_CXXFLAGS)
# combine build flags with cmdline overrides
override CFLAGS := $(MK_CPPFLAGS) $(CPPFLAGS) $(MK_CFLAGS) $(GF_CFLAGS) $(CFLAGS)
BASE_CXXFLAGS := $(MK_CPPFLAGS) $(CPPFLAGS) $(MK_CXXFLAGS) $(CXXFLAGS)
override CXXFLAGS := $(BASE_CXXFLAGS) $(HOST_CXXFLAGS) $(GF_CXXFLAGS)
override NVCCFLAGS := $(MK_NVCCFLAGS) $(NVCCFLAGS)
override LDFLAGS := $(MK_LDFLAGS) $(LDFLAGS)
# identify CUDA host compiler
ifdef LLAMA_CUBLAS
GF_CC := $(NVCC) $(NVCCFLAGS) 2>/dev/null .c -Xcompiler
include scripts/get-flags.mk
CUDA_CXXFLAGS := $(GF_CXXFLAGS)
endif
#
# Print build information
@ -626,7 +628,7 @@ save-load-state: examples/save-load-state/save-load-state.cpp ggml.o llama.o $(C
server: examples/server/server.cpp examples/server/httplib.h examples/server/json.hpp examples/server/index.html.hpp examples/server/index.js.hpp examples/server/completion.js.hpp examples/llava/clip.cpp examples/llava/clip.h common/stb_image.h ggml.o llama.o $(COMMON_DEPS) grammar-parser.o $(OBJS)
$(CXX) $(CXXFLAGS) -Iexamples/server $(filter-out %.h,$(filter-out %.hpp,$^)) -o $@ $(LDFLAGS) $(LWINSOCK2) -Wno-cast-qual
gguf: examples/gguf/gguf.cpp ggml.o llama.o $(OBJS)
gguf: examples/gguf/gguf.cpp ggml.o $(OBJS)
$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
train-text-from-scratch: examples/train-text-from-scratch/train-text-from-scratch.cpp ggml.o llama.o $(COMMON_DEPS) train.o $(OBJS)
@ -653,7 +655,7 @@ beam-search: examples/beam-search/beam-search.cpp ggml.o llama.o $(COMMON_DEPS)
finetune: examples/finetune/finetune.cpp ggml.o llama.o $(COMMON_DEPS) train.o $(OBJS)
$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
export-lora: examples/export-lora/export-lora.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS)
export-lora: examples/export-lora/export-lora.cpp ggml.o common/common.h $(OBJS)
$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
speculative: examples/speculative/speculative.cpp ggml.o llama.o $(COMMON_DEPS) grammar-parser.o $(OBJS)
@ -662,6 +664,12 @@ speculative: examples/speculative/speculative.cpp ggml.o llama.o $(COMMON_DEPS)
parallel: examples/parallel/parallel.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS)
$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
lookahead: examples/lookahead/lookahead.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS)
$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
lookup: examples/lookup/lookup.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS)
$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
ifdef LLAMA_METAL
metal: examples/metal/metal.cpp ggml.o $(OBJS)
$(CXX) $(CXXFLAGS) $^ -o $@ $(LDFLAGS)
@ -703,34 +711,34 @@ vdot: pocs/vdot/vdot.cpp ggml.o $(OBJS)
q8dot: pocs/vdot/q8dot.cpp ggml.o $(OBJS)
$(CXX) $(CXXFLAGS) $^ -o $@ $(LDFLAGS)
tests/test-llama-grammar: tests/test-llama-grammar.cpp ggml.o $(COMMON_DEPS) grammar-parser.o $(OBJS)
tests/test-llama-grammar: tests/test-llama-grammar.cpp ggml.o grammar-parser.o $(OBJS)
$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
tests/test-grammar-parser: tests/test-grammar-parser.cpp ggml.o llama.o $(COMMON_DEPS) grammar-parser.o $(OBJS)
tests/test-grammar-parser: tests/test-grammar-parser.cpp ggml.o llama.o grammar-parser.o $(OBJS)
$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
tests/test-double-float: tests/test-double-float.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS)
tests/test-double-float: tests/test-double-float.cpp ggml.o $(OBJS)
$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
tests/test-grad0: tests/test-grad0.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS)
tests/test-grad0: tests/test-grad0.cpp ggml.o $(OBJS)
$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
tests/test-opt: tests/test-opt.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS)
tests/test-opt: tests/test-opt.cpp ggml.o $(OBJS)
$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
tests/test-quantize-fns: tests/test-quantize-fns.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS)
tests/test-quantize-fns: tests/test-quantize-fns.cpp ggml.o $(OBJS)
$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
tests/test-quantize-perf: tests/test-quantize-perf.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS)
tests/test-quantize-perf: tests/test-quantize-perf.cpp ggml.o $(OBJS)
$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
tests/test-sampling: tests/test-sampling.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS)
tests/test-sampling: tests/test-sampling.cpp ggml.o llama.o $(OBJS)
$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
tests/test-tokenizer-0-falcon: tests/test-tokenizer-0-falcon.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS)
tests/test-tokenizer-0-falcon: tests/test-tokenizer-0-falcon.cpp ggml.o llama.o $(COMMON_DEPS) console.o $(OBJS)
$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
tests/test-tokenizer-0-llama: tests/test-tokenizer-0-llama.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS)
tests/test-tokenizer-0-llama: tests/test-tokenizer-0-llama.cpp ggml.o llama.o $(COMMON_DEPS) console.o $(OBJS)
$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
tests/test-tokenizer-0-deepseek-coder: tests/test-tokenizer-0-deepseek-coder.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS)
@ -739,11 +747,17 @@ tests/test-tokenizer-0-deepseek-coder: tests/test-tokenizer-0-deepseek-coder.cpp
tests/test-tokenizer-0-deepseek-llm: tests/test-tokenizer-0-deepseek-llm.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS)
$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
tests/test-tokenizer-1-bpe: tests/test-tokenizer-1-bpe.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS)
tests/test-tokenizer-1-bpe: tests/test-tokenizer-1-bpe.cpp ggml.o llama.o $(COMMON_DEPS) console.o $(OBJS)
$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
tests/test-tokenizer-1-llama: tests/test-tokenizer-1-llama.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS)
tests/test-tokenizer-1-llama: tests/test-tokenizer-1-llama.cpp ggml.o llama.o $(COMMON_DEPS) console.o $(OBJS)
$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
tests/test-rope: tests/test-rope.cpp ggml.o $(OBJS)
$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
tests/test-c.o: tests/test-c.c llama.h
$(CC) $(CFLAGS) -c $(filter-out %.h,$^) -o $@
tests/test-backend-ops: tests/test-backend-ops.cpp ggml.o $(OBJS)
$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)

46
Package.swift
View file

@ -2,33 +2,14 @@
import PackageDescription
#if arch(arm) || arch(arm64)
let platforms: [SupportedPlatform]? = [
.macOS(.v12),
.iOS(.v14),
.watchOS(.v4),
.tvOS(.v14)
]
let exclude: [String] = []
let resources: [Resource] = [
.process("ggml-metal.metal")
]
let additionalSources: [String] = ["ggml-metal.m"]
let additionalSettings: [CSetting] = [
.unsafeFlags(["-fno-objc-arc"]),
.define("GGML_USE_METAL")
]
#else
let platforms: [SupportedPlatform]? = nil
let exclude: [String] = ["ggml-metal.metal"]
let resources: [Resource] = []
let additionalSources: [String] = []
let additionalSettings: [CSetting] = []
#endif
let package = Package(
name: "llama",
platforms: platforms,
platforms: [
.macOS(.v12),
.iOS(.v14),
.watchOS(.v4),
.tvOS(.v14)
],
products: [
.library(name: "llama", targets: ["llama"]),
],
@ -36,25 +17,30 @@ let package = Package(
.target(
name: "llama",
path: ".",
exclude: exclude,
exclude: [],
sources: [
"ggml.c",
"llama.cpp",
"ggml-alloc.c",
"ggml-backend.c",
"ggml-quants.c",
] + additionalSources,
resources: resources,
"ggml-metal.m",
],
resources: [
.process("ggml-metal.metal")
],
publicHeadersPath: "spm-headers",
cSettings: [
.unsafeFlags(["-Wno-shorten-64-to-32", "-O3", "-DNDEBUG"]),
.define("GGML_USE_ACCELERATE")
.define("GGML_USE_ACCELERATE"),
.unsafeFlags(["-fno-objc-arc"]),
.define("GGML_USE_METAL"),
// NOTE: NEW_LAPACK will required iOS version 16.4+
// We should consider add this in the future when we drop support for iOS 14
// (ref: ref: https://developer.apple.com/documentation/accelerate/1513264-cblas_sgemm?language=objc)
// .define("ACCELERATE_NEW_LAPACK"),
// .define("ACCELERATE_LAPACK_ILP64")
] + additionalSettings,
],
linkerSettings: [
.linkedFramework("Accelerate")
]

54
README.md
View file

@ -10,7 +10,11 @@ Inference of [LLaMA](https://arxiv.org/abs/2302.13971) model in pure C/C++
### Hot topics
- *No hot topics atm. Open to suggestions about what is hot today*
- Collecting Apple Silicon performance stats:
- M-series: https://github.com/ggerganov/llama.cpp/discussions/4167
- A-series: https://github.com/ggerganov/llama.cpp/discussions/4508
- Added Mixtral support: https://github.com/ggerganov/llama.cpp/pull/4406
- Looking for contributions to improve and maintain the `server` example: https://github.com/ggerganov/llama.cpp/issues/4216
----
@ -93,7 +97,19 @@ as the main playground for developing new features for the [ggml](https://github
- [X] [Persimmon 8B](https://github.com/ggerganov/llama.cpp/pull/3410)
- [X] [MPT](https://github.com/ggerganov/llama.cpp/pull/3417)
- [X] [Bloom](https://github.com/ggerganov/llama.cpp/pull/3553)
- [x] [Yi models](https://huggingface.co/models?search=01-ai/Yi)
- [X] [StableLM-3b-4e1t](https://github.com/ggerganov/llama.cpp/pull/3586)
- [x] [Deepseek models](https://huggingface.co/models?search=deepseek-ai/deepseek)
- [x] [Qwen models](https://huggingface.co/models?search=Qwen/Qwen)
- [x] [Mixtral MoE](https://huggingface.co/models?search=mistral-ai/Mixtral)
- [x] [PLaMo-13B](https://github.com/ggerganov/llama.cpp/pull/3557)
**Multimodal models:**
- [x] [Llava 1.5 models](https://huggingface.co/collections/liuhaotian/llava-15-653aac15d994e992e2677a7e)
- [x] [Bakllava](https://huggingface.co/models?search=SkunkworksAI/Bakllava)
- [x] [Obsidian](https://huggingface.co/NousResearch/Obsidian-3B-V0.5)
- [x] [ShareGPT4V](https://huggingface.co/models?search=Lin-Chen/ShareGPT4V)
**Bindings:**
@ -108,12 +124,16 @@ as the main playground for developing new features for the [ggml](https://github
- Clojure: [phronmophobic/llama.clj](https://github.com/phronmophobic/llama.clj)
- React Native: [mybigday/llama.rn](https://github.com/mybigday/llama.rn)
- Java: [kherud/java-llama.cpp](https://github.com/kherud/java-llama.cpp)
- Zig: [deins/llama.cpp.zig](https://github.com/Deins/llama.cpp.zig)
**UI:**
- [nat/openplayground](https://github.com/nat/openplayground)
- [oobabooga/text-generation-webui](https://github.com/oobabooga/text-generation-webui)
- [withcatai/catai](https://github.com/withcatai/catai)
- [semperai/amica](https://github.com/semperai/amica)
- [psugihara/FreeChat](https://github.com/psugihara/FreeChat)
- [ptsochantaris/emeltal](https://github.com/ptsochantaris/emeltal)
---
@ -320,7 +340,7 @@ mpirun -hostfile hostfile -n 3 ./main -m ./models/7B/ggml-model-q4_0.gguf -n 128
### BLAS Build
Building the program with BLAS support may lead to some performance improvements in prompt processing using batch sizes higher than 32 (the default is 512). BLAS doesn't affect the normal generation performance. There are currently three different implementations of it:
Building the program with BLAS support may lead to some performance improvements in prompt processing using batch sizes higher than 32 (the default is 512). Support with CPU-only BLAS implementations doesn't affect the normal generation performance. We may see generation performance improvements with GPU-involved BLAS implementations, e.g. cuBLAS, hipBLAS and CLBlast. There are currently several different BLAS implementations available for build and use:
- #### Accelerate Framework:
@ -378,6 +398,9 @@ Building the program with BLAS support may lead to some performance improvements
- #### cuBLAS
This provides BLAS acceleration using the CUDA cores of your Nvidia GPU. Make sure to have the CUDA toolkit installed. You can download it from your Linux distro's package manager (e.g. `apt install nvidia-cuda-toolkit`) or from here: [CUDA Toolkit](https://developer.nvidia.com/cuda-downloads).
For Jetson user, if you have Jetson Orin, you can try this: [Offical Support](https://www.jetson-ai-lab.com/tutorial_text-generation.html). If you are using an old model(nano/TX2), need some additional operations before compiling.
- Using `make`:
```bash
make LLAMA_CUBLAS=1
@ -415,25 +438,34 @@ Building the program with BLAS support may lead to some performance improvements
```bash
make LLAMA_HIPBLAS=1
```
- Using `CMake` for Linux:
- Using `CMake` for Linux (assuming a gfx1030-compatible AMD GPU):
```bash
mkdir build
cd build
CC=/opt/rocm/llvm/bin/clang CXX=/opt/rocm/llvm/bin/clang++ cmake .. -DLLAMA_HIPBLAS=ON
cmake --build .
CC=/opt/rocm/llvm/bin/clang CXX=/opt/rocm/llvm/bin/clang++ \
cmake -H. -Bbuild -DLLAMA_HIPBLAS=ON -DAMDGPU_TARGETS=gfx1030 -DCMAKE_BUILD_TYPE=Release \
&& cmake --build build -- -j 16
```
- Using `CMake` for Windows:
On Linux it is also possible to use unified memory architecture (UMA) to share main memory between the CPU and integrated GPU by setting `-DLLAMA_HIP_UMA=ON"`.
However, this hurts performance for non-integrated GPUs (but enables working with integrated GPUs).
- Using `make` (example for target gfx1030, build with 16 CPU threads):
```bash
make -j16 LLAMA_HIPBLAS=1 LLAMA_HIP_UMA=1 AMDGPU_TARGETS=gxf1030
```
- Using `CMake` for Windows (using x64 Native Tools Command Prompt for VS, and assuming a gfx1100-compatible AMD GPU):
```bash
set PATH=%HIP_PATH%\bin;%PATH%
mkdir build
cd build
cmake -G Ninja -DAMDGPU_TARGETS=gfx1100 -DLLAMA_HIPBLAS=ON -DCMAKE_C_COMPILER=clang -DCMAKE_CXX_COMPILER=clang++ ..
cmake --build .
```
Make sure that `AMDGPU_TARGETS` is set to the GPU arch you want to compile for. The above example uses `gfx1100` that corresponds to Radeon RX 7900XTX/XT/GRE. You can find a list of targets [here](https://llvm.org/docs/AMDGPUUsage.html#processors)
Find your gpu version string by matching the most significant version information from `rocminfo | grep gfx | head -1 | awk '{print $2}'` with the list of processors, e.g. `gfx1035` maps to `gfx1030`.
The environment variable [`HIP_VISIBLE_DEVICES`](https://rocm.docs.amd.com/en/latest/understand/gpu_isolation.html#hip-visible-devices) can be used to specify which GPU(s) will be used.
If your GPU is not officially supported you can use the environment variable [`HSA_OVERRIDE_GFX_VERSION`] set to a similar GPU, for example 10.3.0 on RDNA2 or 11.0.0 on RDNA3.
If your GPU is not officially supported you can use the environment variable [`HSA_OVERRIDE_GFX_VERSION`] set to a similar GPU, for example 10.3.0 on RDNA2 (e.g. gfx1030, gfx1031, or gfx1035) or 11.0.0 on RDNA3.
The following compilation options are also available to tweak performance (yes, they refer to CUDA, not HIP, because it uses the same code as the cuBLAS version above):
| Option | Legal values | Default | Description |
@ -891,7 +923,7 @@ Additionally, there the following images, similar to the above:
- `ghcr.io/ggerganov/llama.cpp:full-rocm`: Same as `full` but compiled with ROCm support. (platforms: `linux/amd64`, `linux/arm64`)
- `ghcr.io/ggerganov/llama.cpp:light-rocm`: Same as `light` but compiled with ROCm support. (platforms: `linux/amd64`, `linux/arm64`)
The GPU enabled images are not currently tested by CI beyond being built. They are not built with any variation from the ones in the Dockerfiles defined in [.devops/](.devops/) and the Gitlab Action defined in [.github/workflows/docker.yml](.github/workflows/docker.yml). If you need different settings (for example, a different CUDA or ROCm library, you'll need to build the images locally for now).
The GPU enabled images are not currently tested by CI beyond being built. They are not built with any variation from the ones in the Dockerfiles defined in [.devops/](.devops/) and the GitHub Action defined in [.github/workflows/docker.yml](.github/workflows/docker.yml). If you need different settings (for example, a different CUDA or ROCm library, you'll need to build the images locally for now).
#### Usage
@ -964,6 +996,8 @@ docker run --gpus all -v /path/to/models:/models local/llama.cpp:light-cuda -m /
- There are no strict rules for the code style, but try to follow the patterns in the code (indentation, spaces, etc.). Vertical alignment makes things more readable and easier to batch edit
- Clean-up any trailing whitespaces, use 4 spaces for indentation, brackets on the same line, `void * ptr`, `int & a`
- See [good first issues](https://github.com/ggerganov/llama.cpp/issues?q=is%3Aissue+is%3Aopen+label%3A%22good+first+issue%22) for tasks suitable for first contributions
- Tensors store data in row-major order. We refer to dimension 0 as columns, 1 as rows, 2 as matrices
- Matrix multiplication is unconventional: [`z = ggml_mul_mat(ctx, x, y)`](https://github.com/ggerganov/llama.cpp/blob/880e352277fc017df4d5794f0c21c44e1eae2b84/ggml.h#L1058-L1064) means `zT = x @ yT`
### Docs

116
awq-py/README.md Normal file
View file

@ -0,0 +1,116 @@
# AWQ: Activation-aware Weight Quantization for LLM - version apply to llamacpp
[[Paper](https://arxiv.org/abs/2306.00978)][[Original Repo](https://github.com/mit-han-lab/llm-awq)][[Easy-to-use Repo](https://github.com/casper-hansen/AutoAWQ)]
**Supported models:**
- [X] LLaMA
- [x] LLaMA 2
- [X] MPT
- [X] Mistral AI v0.1
- [ ] Bloom
- [ ] Mixtral MoE
**TODO:**
- [x] Update version work with both MPT and MPT-AWQ model
- [ ] Add OPT model
- [ ] Add Bloom model
- [ ] Add Mixtral MoE
- [ ] Support w3, w2
## Contents
- [Install](##Install)
- [Convert](##Convert)
- [Quantize](##Quantize)
- [Test](##Test)
- [Benchmark](##Benchmark)
- [Results](##Results)
## Install
Install requirements
```bash
pip install -r requirements.txt
```
Get the pre-computed AWQ search results for multiple model families, including LLaMA, LLaMA2, MPT, OPT
```bash
git clone https://huggingface.co/datasets/mit-han-lab/awq-model-zoo awq_cache
```
## Convert
Example for llama model
```bash
# For llama7b and llama2 models
python convert.py models/llama-7b/ --awq-path awq_cache/llama-7b-w4-g128.pt --outfile models/llama_7b_fp16.gguf
# For mistral and mpt models
python convert-hf-to-gguf.py models/mpt-7b/ --awq-path awq_cache/llama-7b-w4-g128.pt --outfile models/mpt_7b_fp16.gguf
```
## Quantize
```bash
# We only benchmark and confirm the results on q4_0, q4_1, and q2_k types.
./quantize models/llama_7b_fp16.gguf models/llama_7b_q4_0.gguf q4_0
```
## Test
```bash
# For all models.
./build/bin/main -m models/llama_7b_q4_0.gguf -n 128 --prompt "Once upon a time"
```
## Benchmark
The perplexity measurements in table above are done against the `wikitext2` test dataset (https://paperswithcode.com/dataset/wikitext-2), with context length of 512.
```bash
# For llama and llama2, and mistral models.
./perplexity -m models/llama_7b_q4_0.gguf -f datasets/wikitext-2-raw/wiki.test.raw
```
## Results
Results are run on OpenBLAS (CPU) and CuBLAS (GPU) for fair comparison
We use three types of llamacpp quantization methods to work with our version, including q4_0, q4_1, and q2_k
### Llama 7B (Build with OpenBLAS)
| Model | Measure | F16 | Q4_0 | Q4_1 | Q2_K |
|-----------:|--------------|-------:|-------:|-------:|-------:|
|Llama 7B | perplexity | 5.9066 | 6.1214 | 6.0643 | 6.5808 |
|Llama 7B | file size | 12.9G | 3.5G | 3.9G | 2.7G |
|Llama 7B | bits/weight | 16.0 | 4.5 | 5.0 | 2.6 |
|AWQ-LLama 7B| perplexity | 5.9175 | 6.0252 | 5.9987 | 6.3692 |
|AWQ-LLama 7B| file size | 12.9G | 3.5G | 3.9G | 2.7G |
|AWQ-LLama 7B| bits/weight | 16.0 | 4.5 | 5.0 | 2.6 |
### Llama2 7B (Build with CuBLAS)
| Model | Measure | F16 | Q4_0 | Q4_1 | Q2_K |
|------------:|--------------|-------:|-------:|-------:|-------:|
|Llama2 7B | perplexity | 5.8664 | 6.0260 | 6.0656 | 6.4496 |
|Llama2 7B | file size | 12.9G | 3.5G | 3.9G | 2.7G |
|Llama2 7B | bits/weight | 16.0 | 4.5 | 5.0 | 2.6 |
|AWQ-LLama2 7B| perplexity | 5.8801 | 6.0054 | 5.9849 | 6.3650 |
|AWQ-LLama2 7B| file size | 12.9G | 3.5G | 3.9G | 2.7G |
|AWQ-LLama2 7B| bits/weight | 16.0 | 4.5 | 5.0 | 2.6 |
### Mistral 7B v0.1 (Build with CuBLAS)
| Model | Measure | F16 | Q4_0 | Q4_1 | Q2_K |
|-------------:|--------------|-------:|-------:|-------:|-------:|
|Mistral 7B | perplexity | 5.6931 | 5.8202 | 5.8268 | 6.1645 |
|Mistral 7B | file size | 14.5G | 4.1G | 4.5G | 3.1G |
|Mistral 7B | bits/weight | 16.0 | 4.5 | 5.0 | 2.6 |
|AWQ-Mistral 7B| perplexity | 5.6934 | 5.8020 | 5.7691 | 6.0426 |
|AWQ-Mistral 7B| file size | 14.5G | 4.1G | 4.5G | 3.1G |
|AWQ-Mistral 7B| bits/weight | 16.0 | 4.5 | 5.0 | 2.6 |
### MPT 7B (Build with OpenBLAS)
| Model | Measure | F16 | Q4_0 | Q4_1 | Q2_K |
|---------:|--------------|-------:|-------:|-------:|--------:|
|MPT 7B | perplexity | 8.4369 | 8.7956 | 8.6265 | 11.4913 |
|MPT 7B | file size | 13.7G | 3.9G | 4.3G | 2.8G |
|MPT 7B | bits/weight | 16.0 | 4.5 | 5.0 | 2.6 |
|AWQ-MPT 7B| perplexity | 8.4944 | 8.7053 | 8.6750 | 10.2873|
|AWQ-MPT 7B| file size | 13.7G | 3.9G | 4.3G | 2.8G |
|AWQ-MPT 7B| bits/weight | 16.0 | 4.5 | 5.0 | 2.6 |

254
awq-py/awq/apply_awq.py Normal file
View file

@ -0,0 +1,254 @@
"""
Implements the AWQ for llama.cpp use cases.
Original paper: https://arxiv.org/abs/2306.00978
This code is based on versions of the AWQ implementation found in the following repositories:
* https://github.com/mit-han-lab/llm-awq
* https://github.com/casper-hansen/AutoAWQ
"""
import os
import torch
import torch.nn as nn
from transformers import AutoModelForCausalLM, AutoConfig
from transformers.models.bloom.modeling_bloom import BloomGelu
from transformers.models.llama.modeling_llama import LlamaRMSNorm
from transformers.activations import GELUActivation
class ScaledActivation(nn.Module):
"""
ScaledActivation module wraps an existing activation function and applies a
scale factor to its output.
Args:
module (nn.Module): The activation function to be scaled.
scales (torch.Tensor): A tensor of size (num_features,) containing the initial
scale factors for each feature.
Returns:
torch.Tensor: The scaled output of the activation function.
"""
def __init__(self, module, scales):
super().__init__()
self.act = module
self.scales = nn.Parameter(scales.data)
def forward(self, x):
return self.act(x) / self.scales.view(1, 1, -1).to(x.device)
def set_op_by_name(layer, name, new_module):
"""
Set the new module for given module's name.
Args:
layer (nn.Module): The layer in which to replace the submodule.
name (str): The path to the submodule to be replaced, using dot notation
to access nested modules.
new_module (nn.Module): The new module to replace the existing one.
"""
levels = name.split(".")
if len(levels) > 1:
mod_ = layer
for l_idx in range(len(levels) - 1):
if levels[l_idx].isdigit():
mod_ = mod_[int(levels[l_idx])]
else:
mod_ = getattr(mod_, levels[l_idx])
setattr(mod_, levels[-1], new_module)
else:
setattr(layer, name, new_module)
def get_op_by_name(module, op_name):
"""
Retrieves a submodule within a given layer based on its name.
Args:
module (nn.Module): The layer containing the submodule to find.
op_name (str): The name of the submodule.
Returns:
nn.Module: The requested submodule found within the given layer.
Raises:
ValueError: If the specified submodule cannot be found within the layer.
"""
for name, m in module.named_modules():
if name == op_name:
return m
raise ValueError(f"Cannot find op {op_name} in module {module}")
@torch.no_grad()
def scale_ln_fcs(ln, fcs, scales):
"""
Scales the weights of a LayerNorm and a list of fully-connected layers proportionally.
Args:
ln (nn.LayerNorm): The LayerNorm module to be scaled.
fcs (List[nn.Linear]): A list of fully-connected layers to be scaled.
scales (torch.Tensor): A 1D tensor of size (num_features,).
"""
if not isinstance(fcs, list):
fcs = [fcs]
scales = scales.to(ln.weight.device)
ln.weight.div_(scales)
if hasattr(ln, "bias") and ln.bias is not None:
ln.bias.div_(scales)
for fc in fcs:
fc.weight.mul_(scales.view(1, -1))
for p in ln.parameters():
assert torch.isnan(p).sum() == 0
for fc in fcs:
for p in fc.parameters():
assert torch.isnan(p).sum() == 0
@torch.no_grad()
def scale_fc_fc(fc1, fc2, scales):
"""
Scales the weights of two fully-connected layers in a specific pattern.
Args:
fc1 (nn.Linear): The first fully-connected layer to be scaled.
fc2 (nn.Linear): The second fully-connected layer to be scaled.
scales (torch.Tensor): A 1D tensor of size (num_features,).
"""
assert isinstance(fc1, nn.Linear)
assert isinstance(fc2, nn.Linear)
scales = scales.to(fc1.weight.device)
fc1.weight[-scales.size(0):].div_(scales.view(-1, 1))
if fc1.bias is not None:
fc1.bias.div_(scales.view(-1))
fc2.weight.mul_(scales.view(1, -1))
for p in fc1.parameters():
assert torch.isnan(p).sum() == 0
for p in fc2.parameters():
assert torch.isnan(p).sum() == 0
@torch.no_grad()
def scale_gelu_fc(gelu, fc, scales):
"""
Scales the weight of a GELU activation and a fully-connected layer proportionally.
Args:
gelu (Union[nn.GELU, BloomGelu, GELUActivation]): The GELU activation module to be scaled.
fc (nn.Linear): The fully-connected layer to be scaled.
scales (torch.Tensor): A 1D tensor of size (num_features,).
Raises:
TypeError: If the `gelu` module is not of type `nn.GELU`, `BloomGelu`, or `GELUActivation`.
TypeError: If the `fc` module is not of type `nn.Linear`.
"""
assert isinstance(gelu, (nn.GELU, BloomGelu, GELUActivation))
assert isinstance(fc, nn.Linear)
fc.weight.mul_(scales.view(1, -1).to(fc.weight.device))
for p in fc.parameters():
assert torch.isnan(p).sum() == 0
def apply_scale(module, scales_list, input_feat_dict=None):
"""
Applies different scaling strategies to layers based on their type and hierarchy within a given module.
Args:
module (nn.Module): The module containing the layers to be scaled.
scales_list (List[Tuple[str, List[str], torch.Tensor]]): A list of tuples containing:
* prev_op_name (str): The name of the preceding operation or module,
relative to which the layers to be scaled are located.
* layer_names (List[str]): A list of names of the layers to be scaled, relative to the preceding operation.
* scales (torch.Tensor): A 1D tensor of size (num_features,) containing the scaling factors for each feature.
input_feat_dict (Optional[Dict[str, torch.Tensor]]): A dictionary mapping layer names to their corresponding
input features (optional).
"""
for prev_op_name, layer_names, scales in scales_list:
prev_op = get_op_by_name(module, prev_op_name)
layers = [get_op_by_name(module, name) for name in layer_names]
prev_op.cuda()
for layer in layers:
layer.cuda()
scales.cuda()
if isinstance(prev_op, nn.Linear):
assert len(layers) == 1
scale_fc_fc(prev_op, layers[0], scales)
elif isinstance(prev_op, (nn.LayerNorm, LlamaRMSNorm)) or "rmsnorm" in str(prev_op.__class__).lower():
scale_ln_fcs(prev_op, layers, scales)
elif isinstance(prev_op, (nn.GELU, BloomGelu, GELUActivation)):
new_module = ScaledActivation(prev_op, scales)
set_op_by_name(module, prev_op_name, new_module)
scale_gelu_fc(prev_op, layers[0], scales)
else:
raise NotImplementedError(f"prev_op {type(prev_op)} not supported yet!")
# apply the scaling to input feat if given; prepare it for clipping
if input_feat_dict is not None:
for layer_name in layer_names:
inp = input_feat_dict[layer_name]
inp.div_(scales.view(1, -1).to(inp.device))
prev_op.cpu()
for layer in layers:
layer.cpu()
scales.cpu()
@torch.no_grad()
def apply_clip(module, clip_list):
"""
Applies element-wise clipping to the weight of a specific layer within a given module.
Args:
module (nn.Module): The module containing the layer to be clipped.
clip_list (List[Tuple[str, torch.Tensor]]): A list of tuples containing:
* name (str): The name of the layer to be clipped, relative to the root of the module.
* max_val (torch.Tensor): A 1D or 2D tensor defining the upper bound for each element of the layer's weight.
"""
for name, max_val in clip_list:
layer = get_op_by_name(module, name)
layer.cuda()
max_val = max_val.to(layer.weight.device)
org_shape = layer.weight.shape
layer.weight.data = layer.weight.data.reshape(*max_val.shape[:2], -1)
layer.weight.data = torch.clamp(layer.weight.data, -max_val, max_val)
layer.weight.data = layer.weight.data.reshape(org_shape)
layer.cpu()
def add_scale_weights(model_path, scale_path, tmp_path):
"""
Adds pre-computed Activation Weight Quantization (AWQ) results to a model,
including scaling factors and clipping bounds.
Args:
model_path (str): Path to the pre-trained model to be equipped with AWQ.
scale_path (str): Path to the AWQ scale factors (.pt file).
tmp_path (str): Path to the temporary directory where the equipped model will be saved.
"""
config = AutoConfig.from_pretrained(model_path, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
model_path, config=config, trust_remote_code=True
)
model.eval()
awq_results = torch.load(str(scale_path), map_location="cpu")
apply_scale(model, awq_results["scale"])
apply_clip(model, awq_results["clip"])
model.save_pretrained(str(tmp_path))
os.system(f"cp {str(model_path)}/tokenizer* {str(tmp_path)}")

2
awq-py/requirements.txt Normal file
View file

@ -0,0 +1,2 @@
torch>=2.0.0
transformers>=4.32.0

9
common/CMakeLists.txt
View file

@ -11,7 +11,12 @@ if(EXISTS "${CMAKE_CURRENT_SOURCE_DIR}/../.git")
if(NOT IS_DIRECTORY "${GIT_DIR}")
file(READ ${GIT_DIR} REAL_GIT_DIR_LINK)
string(REGEX REPLACE "gitdir: (.*)\n$" "\\1" REAL_GIT_DIR ${REAL_GIT_DIR_LINK})
set(GIT_DIR "${CMAKE_CURRENT_SOURCE_DIR}/../${REAL_GIT_DIR}")
string(FIND "${REAL_GIT_DIR}" "/" SLASH_POS)
if (SLASH_POS EQUAL 0)
set(GIT_DIR "${REAL_GIT_DIR}")
else()
set(GIT_DIR "${CMAKE_CURRENT_SOURCE_DIR}/../${REAL_GIT_DIR}")
endif()
endif()
set(GIT_INDEX "${GIT_DIR}/index")
@ -26,7 +31,7 @@ add_custom_command(
COMMENT "Generating build details from Git"
COMMAND ${CMAKE_COMMAND} -DMSVC=${MSVC} -DCMAKE_C_COMPILER_VERSION=${CMAKE_C_COMPILER_VERSION}
-DCMAKE_C_COMPILER_ID=${CMAKE_C_COMPILER_ID} -DCMAKE_VS_PLATFORM_NAME=${CMAKE_VS_PLATFORM_NAME}
-DCMAKE_C_COMPILER=${CMAKE_C_COMPILER} -P "${CMAKE_CURRENT_SOURCE_DIR}/../scripts/build-info.cmake"
-DCMAKE_C_COMPILER=${CMAKE_C_COMPILER} -P "${CMAKE_CURRENT_SOURCE_DIR}/../scripts/gen-build-info-cpp.cmake"
WORKING_DIRECTORY "${CMAKE_CURRENT_SOURCE_DIR}/.."
DEPENDS "${CMAKE_CURRENT_SOURCE_DIR}/build-info.cpp.in" ${GIT_INDEX}
VERBATIM

242
common/common.cpp
View file

@ -12,6 +12,7 @@
#include <regex>
#include <sstream>
#include <string>
#include <unordered_map>
#include <unordered_set>
#include <vector>
#include <cinttypes>
@ -277,8 +278,18 @@ bool gpt_params_parse_ex(int argc, char ** argv, gpt_params & params) {
break;
}
params.yarn_beta_slow = std::stof(argv[i]);
} else if (arg == "--memory-f32") {
params.memory_f16 = false;
} else if (arg == "--samplers") {
if (++i >= argc) {
invalid_param = true;
break;
}
sparams.samplers_sequence = parse_samplers_input(argv[i]);
} else if (arg == "--sampling-seq") {
if (++i >= argc) {
invalid_param = true;
break;
}
sparams.samplers_sequence = argv[i];
} else if (arg == "--top-p") {
if (++i >= argc) {
invalid_param = true;
@ -495,6 +506,14 @@ bool gpt_params_parse_ex(int argc, char ** argv, gpt_params & params) {
params.chatml = true;
} else if (arg == "--infill") {
params.infill = true;
} else if (arg == "-dkvc" || arg == "--dump-kv-cache") {
params.dump_kv_cache = true;
} else if (arg == "-nkvo" || arg == "--no-kv-offload") {
params.no_kv_offload = true;
} else if (arg == "-ctk" || arg == "--cache-type-k") {
params.cache_type_k = argv[++i];
} else if (arg == "-ctv" || arg == "--cache-type-v") {
params.cache_type_v = argv[++i];
} else if (arg == "--multiline-input") {
params.multiline_input = true;
} else if (arg == "--simple-io") {
@ -637,6 +656,10 @@ bool gpt_params_parse_ex(int argc, char ** argv, gpt_params & params) {
} else if (arg == "-h" || arg == "--help") {
return false;
} else if (arg == "--version") {
fprintf(stderr, "version: %d (%s)\n", LLAMA_BUILD_NUMBER, LLAMA_COMMIT);
fprintf(stderr, "built with %s for %s\n", LLAMA_COMPILER, LLAMA_BUILD_TARGET);
exit(0);
} else if (arg == "--random-prompt") {
params.random_prompt = true;
} else if (arg == "--in-prefix-bos") {
@ -675,6 +698,47 @@ bool gpt_params_parse_ex(int argc, char ** argv, gpt_params & params) {
std::istreambuf_iterator<char>(),
std::back_inserter(sparams.grammar)
);
} else if (arg == "--override-kv") {
if (++i >= argc) {
invalid_param = true;
break;
}
char * sep = strchr(argv[i], '=');
if (sep == nullptr || sep - argv[i] >= 128) {
fprintf(stderr, "error: Malformed KV override: %s\n", argv[i]);
invalid_param = true;
break;
}
struct llama_model_kv_override kvo;
std::strncpy(kvo.key, argv[i], sep - argv[i]);
kvo.key[sep - argv[i]] = 0;
sep++;
if (strncmp(sep, "int:", 4) == 0) {
sep += 4;
kvo.tag = LLAMA_KV_OVERRIDE_INT;
kvo.int_value = std::atol(sep);
} else if (strncmp(sep, "float:", 6) == 0) {
sep += 6;
kvo.tag = LLAMA_KV_OVERRIDE_FLOAT;
kvo.float_value = std::atof(sep);
} else if (strncmp(sep, "bool:", 5) == 0) {
sep += 5;
kvo.tag = LLAMA_KV_OVERRIDE_BOOL;
if (std::strcmp(sep, "true") == 0) {
kvo.bool_value = true;
} else if (std::strcmp(sep, "false") == 0) {
kvo.bool_value = false;
} else {
fprintf(stderr, "error: Invalid boolean value for KV override: %s\n", argv[i]);
invalid_param = true;
break;
}
} else {
fprintf(stderr, "error: Invalid type for KV override: %s\n", argv[i]);
invalid_param = true;
break;
}
params.kv_overrides.push_back(kvo);
#ifndef LOG_DISABLE_LOGS
// Parse args for logging parameters
} else if ( log_param_single_parse( argv[i] ) ) {
@ -718,6 +782,11 @@ bool gpt_params_parse_ex(int argc, char ** argv, gpt_params & params) {
}
}
if (!params.kv_overrides.empty()) {
params.kv_overrides.emplace_back(llama_model_kv_override());
params.kv_overrides.back().key[0] = 0;
}
return true;
}
@ -729,6 +798,7 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) {
printf("\n");
printf("options:\n");
printf(" -h, --help show this help message and exit\n");
printf(" --version show version and build info\n");
printf(" -i, --interactive run in interactive mode\n");
printf(" --interactive-first run in interactive mode and wait for input right away\n");
printf(" -ins, --instruct run in instruction mode (use with Alpaca models)\n");
@ -758,6 +828,8 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) {
printf(" -n N, --n-predict N number of tokens to predict (default: %d, -1 = infinity, -2 = until context filled)\n", params.n_predict);
printf(" -c N, --ctx-size N size of the prompt context (default: %d, 0 = loaded from model)\n", params.n_ctx);
printf(" -b N, --batch-size N batch size for prompt processing (default: %d)\n", params.n_batch);
printf(" --samplers samplers that will be used for generation in the order, separated by \';\', for example: \"top_k;tfs;typical;top_p;min_p;temp\"\n");
printf(" --sampling-seq simplified sequence for samplers that will be used (default: %s)\n", sparams.samplers_sequence.c_str());
printf(" --top-k N top-k sampling (default: %d, 0 = disabled)\n", sparams.top_k);
printf(" --top-p N top-p sampling (default: %.1f, 1.0 = disabled)\n", (double)sparams.top_p);
printf(" --min-p N min-p sampling (default: %.1f, 0.0 = disabled)\n", (double)sparams.min_p);
@ -795,8 +867,6 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) {
printf(" --yarn-beta-fast N YaRN: low correction dim or beta (default: %.1f)\n", params.yarn_beta_fast);
printf(" --ignore-eos ignore end of stream token and continue generating (implies --logit-bias 2-inf)\n");
printf(" --no-penalize-nl do not penalize newline token\n");
printf(" --memory-f32 use f32 instead of f16 for memory key+value (default: disabled)\n");
printf(" not recommended: doubles context memory required and no measurable increase in quality\n");
printf(" --temp N temperature (default: %.1f)\n", (double)sparams.temp);
printf(" --logits-all return logits for all tokens in the batch (default: disabled)\n");
printf(" --hellaswag compute HellaSwag score over random tasks from datafile supplied with -f\n");
@ -835,6 +905,14 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) {
#endif // GGML_USE_CUBLAS
#endif
printf(" --verbose-prompt print prompt before generation\n");
printf(" -dkvc, --dump-kv-cache\n");
printf(" verbose print of the KV cache\n");
printf(" -nkvo, --no-kv-offload\n");
printf(" disable KV offload\n");
printf(" -ctk TYPE, --cache-type-k TYPE\n");
printf(" KV cache data type for K (default: %s)\n", params.cache_type_k.c_str());
printf(" -ctv TYPE, --cache-type-v TYPE\n");
printf(" KV cache data type for V (default: %s)\n", params.cache_type_v.c_str());
printf(" --simple-io use basic IO for better compatibility in subprocesses and limited consoles\n");
printf(" --lora FNAME apply LoRA adapter (implies --no-mmap)\n");
printf(" --lora-scaled FNAME S apply LoRA adapter with user defined scaling S (implies --no-mmap)\n");
@ -842,9 +920,12 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) {
printf(" -m FNAME, --model FNAME\n");
printf(" model path (default: %s)\n", params.model.c_str());
printf(" -md FNAME, --model-draft FNAME\n");
printf(" draft model for speculative decoding (default: %s)\n", params.model.c_str());
printf(" draft model for speculative decoding\n");
printf(" -ld LOGDIR, --logdir LOGDIR\n");
printf(" path under which to save YAML logs (no logging if unset)\n");
printf(" --override-kv KEY=TYPE:VALUE\n");
printf(" advanced option to override model metadata by key. may be specified multiple times.\n");
printf(" types: int, float, bool. example: --override-kv tokenizer.ggml.add_bos_token=bool:false\n");
printf("\n");
#ifndef LOG_DISABLE_LOGS
log_print_usage();
@ -881,6 +962,48 @@ std::string gpt_random_prompt(std::mt19937 & rng) {
GGML_UNREACHABLE();
}
//
// String parsing
//
std::string parse_samplers_input(std::string input) {
std::string output = "";
// since samplers names are written multiple ways
// make it ready for both system names and input names
std::unordered_map<std::string, char> samplers_symbols {
{"top_k", 'k'},
{"top-k", 'k'},
{"top_p", 'p'},
{"top-p", 'p'},
{"nucleus", 'p'},
{"typical_p", 'y'},
{"typical-p", 'y'},
{"typical", 'y'},
{"min_p", 'm'},
{"min-p", 'm'},
{"tfs_z", 'f'},
{"tfs-z", 'f'},
{"tfs", 'f'},
{"temp", 't'},
{"temperature",'t'}
};
// expected format example: "temp;top_k;tfs_z;typical_p;top_p;min_p"
size_t separator = input.find(';');
while (separator != input.npos) {
std::string name = input.substr(0,separator);
input = input.substr(separator+1);
separator = input.find(';');
if (samplers_symbols.find(name) != samplers_symbols.end()) {
output += samplers_symbols[name];
}
}
if (samplers_symbols.find(input) != samplers_symbols.end()) {
output += samplers_symbols[input];
}
return output;
}
//
// Model utils
//
@ -895,10 +1018,39 @@ struct llama_model_params llama_model_params_from_gpt_params(const gpt_params &
mparams.tensor_split = params.tensor_split;
mparams.use_mmap = params.use_mmap;
mparams.use_mlock = params.use_mlock;
if (params.kv_overrides.empty()) {
mparams.kv_overrides = NULL;
} else {
GGML_ASSERT(params.kv_overrides.back().key[0] == 0 && "KV overrides not terminated with empty key");
mparams.kv_overrides = params.kv_overrides.data();
}
return mparams;
}
static ggml_type kv_cache_type_from_str(const std::string & s) {
if (s == "f16") {
return GGML_TYPE_F16;
}
if (s == "q8_0") {
return GGML_TYPE_Q8_0;
}
if (s == "q4_0") {
return GGML_TYPE_Q4_0;
}
if (s == "q4_1") {
return GGML_TYPE_Q4_1;
}
if (s == "q5_0") {
return GGML_TYPE_Q5_0;
}
if (s == "q5_1") {
return GGML_TYPE_Q5_1;
}
throw std::runtime_error("Invalid cache type: " + s);
}
struct llama_context_params llama_context_params_from_gpt_params(const gpt_params & params) {
auto cparams = llama_context_default_params();
@ -908,7 +1060,6 @@ struct llama_context_params llama_context_params_from_gpt_params(const gpt_param
cparams.n_threads_batch = params.n_threads_batch == -1 ? params.n_threads : params.n_threads_batch;
cparams.mul_mat_q = params.mul_mat_q;
cparams.seed = params.seed;
cparams.f16_kv = params.memory_f16;
cparams.logits_all = params.logits_all;
cparams.embedding = params.embedding;
cparams.rope_scaling_type = params.rope_scaling_type;
@ -919,6 +1070,10 @@ struct llama_context_params llama_context_params_from_gpt_params(const gpt_param
cparams.yarn_beta_fast = params.yarn_beta_fast;
cparams.yarn_beta_slow = params.yarn_beta_slow;
cparams.yarn_orig_ctx = params.yarn_orig_ctx;
cparams.offload_kqv = !params.no_kv_offload;
cparams.type_k = kv_cache_type_from_str(params.cache_type_k);
cparams.type_v = kv_cache_type_from_str(params.cache_type_v);
return cparams;
}
@ -1331,7 +1486,6 @@ void dump_non_result_info_yaml(FILE * stream, const gpt_params & params, const l
}
fprintf(stream, "lora_base: %s\n", params.lora_base.c_str());
fprintf(stream, "main_gpu: %d # default: 0\n", params.main_gpu);
fprintf(stream, "memory_f32: %s # default: false\n", !params.memory_f16 ? "true" : "false");
fprintf(stream, "mirostat: %d # default: 0 (disabled)\n", sparams.mirostat);
fprintf(stream, "mirostat_ent: %f # default: 5.0\n", sparams.mirostat_tau);
fprintf(stream, "mirostat_lr: %f # default: 0.1\n", sparams.mirostat_eta);
@ -1386,3 +1540,77 @@ void dump_non_result_info_yaml(FILE * stream, const gpt_params & params, const l
fprintf(stream, "typical_p: %f # default: 1.0\n", sparams.typical_p);
fprintf(stream, "verbose_prompt: %s # default: false\n", params.verbose_prompt ? "true" : "false");
}
//
// KV cache utils
//
void dump_kv_cache_view(const llama_kv_cache_view & view, int row_size) {
static const char slot_chars[] = ".123456789ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz+";
printf("=== Dumping KV cache. total cells %d, max sequences per cell %d, populated cells %d, total tokens in cache %d, largest empty slot=%d @ %d",
view.n_cells, view.n_max_seq, view.used_cells, view.token_count, view.max_contiguous, view.max_contiguous_idx);
llama_kv_cache_view_cell * c_curr = view.cells;
llama_seq_id * cs_curr = view.cells_sequences;
for (int i = 0; i < view.n_cells; i++, c_curr++, cs_curr += view.n_max_seq) {
if (i % row_size == 0) {
printf("\n%5d: ", i);
}
int seq_count = 0;
for (int j = 0; j < view.n_max_seq; j++) {
if (cs_curr[j] >= 0) { seq_count++; }
}
putchar(slot_chars[std::min(sizeof(slot_chars) - 2, size_t(seq_count))]);
}
printf("\n=== Done dumping\n");
}
void dump_kv_cache_view_seqs(const llama_kv_cache_view & view, int row_size) {
static const char slot_chars[] = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz";
printf("=== Dumping KV cache. total cells %d, max sequences per cell %d, populated cells %d, total tokens in cache %d, largest empty slot=%d @ %d\n",
view.n_cells, view.n_max_seq, view.used_cells, view.token_count, view.max_contiguous, view.max_contiguous_idx);
std::unordered_map<llama_seq_id, size_t> seqs;
llama_kv_cache_view_cell * c_curr = view.cells;
llama_seq_id * cs_curr = view.cells_sequences;
for (int i = 0; i < view.n_cells; i++, c_curr++, cs_curr += view.n_max_seq) {
for (int j = 0; j < view.n_max_seq; j++) {
if (cs_curr[j] < 0) { continue; }
if (seqs.find(cs_curr[j]) == seqs.end()) {
if (seqs.size() + 1 >= sizeof(slot_chars)) { break; }
seqs[cs_curr[j]] = seqs.size();
}
}
if (seqs.size() + 1 >= sizeof(slot_chars)) { break; }
}
printf("=== Sequence legend: ");
for (const auto & it : seqs) {
printf("%zu=%d, ", it.second, it.first);
}
printf("'+'=other sequence ids");
c_curr = view.cells;
cs_curr = view.cells_sequences;
for (int i = 0; i < view.n_cells; i++, c_curr++, cs_curr += view.n_max_seq) {
if (i % row_size == 0) {
printf("\n%5d: ", i);
}
for (int j = 0; j < view.n_max_seq; j++) {
if (cs_curr[j] >= 0) {
const auto & it = seqs.find(cs_curr[j]);
putchar(it != seqs.end() ? int(slot_chars[it->second]) : '+');
} else {
putchar('.');
}
}
putchar(' ');
}
printf("\n=== Done dumping\n");
}

29
common/common.h
View file

@ -51,7 +51,7 @@ struct gpt_params {
int32_t n_ctx = 512; // context size
int32_t n_batch = 512; // batch size for prompt processing (must be >=32 to use BLAS)
int32_t n_keep = 0; // number of tokens to keep from initial prompt
int32_t n_draft = 16; // number of tokens to draft during speculative decoding
int32_t n_draft = 8; // number of tokens to draft during speculative decoding
int32_t n_chunks = -1; // max number of chunks to process (-1 = unlimited)
int32_t n_parallel = 1; // number of parallel sequences to decode
int32_t n_sequences = 1; // number of sequences to decode
@ -86,6 +86,8 @@ struct gpt_params {
std::vector<std::string> antiprompt; // string upon seeing which more user input is prompted
std::string logdir = ""; // directory in which to save YAML log files
std::vector<llama_model_kv_override> kv_overrides;
// TODO: avoid tuple, use struct
std::vector<std::tuple<std::string, float>> lora_adapter; // lora adapter path with user defined scale
std::string lora_base = ""; // base model path for the lora adapter
@ -98,7 +100,6 @@ struct gpt_params {
size_t hellaswag_tasks = 400; // number of tasks to use when computing the HellaSwag score
bool mul_mat_q = true; // if true, use mul_mat_q kernels instead of cuBLAS
bool memory_f16 = true; // use f16 instead of f32 for memory kv
bool random_prompt = false; // do not randomize prompt if none provided
bool use_color = false; // use color to distinguish generations and inputs
bool interactive = false; // interactive mode
@ -122,10 +123,15 @@ struct gpt_params {
bool numa = false; // attempt optimizations that help on some NUMA systems
bool verbose_prompt = false; // print prompt tokens before generation
bool infill = false; // use infill mode
bool dump_kv_cache = false; // dump the KV cache contents for debugging purposes
bool no_kv_offload = false; // disable KV offloading
std::string cache_type_k = "f16"; // KV cache data type for the K
std::string cache_type_v = "f16"; // KV cache data type for the V
// multimodal models (see examples/llava)
std::string mmproj = ""; // path to multimodal projector
std::string image = ""; // path to an image file
std::string image = ""; // path to an image file
};
bool gpt_params_parse_ex(int argc, char ** argv, gpt_params & params);
@ -140,6 +146,12 @@ std::string gpt_random_prompt(std::mt19937 & rng);
void process_escapes(std::string& input);
//
// String parsing
//
std::string parse_samplers_input(std::string input);
//
// Model utils
//
@ -218,3 +230,14 @@ std::string get_sortable_timestamp();
void dump_non_result_info_yaml(
FILE * stream, const gpt_params & params, const llama_context * lctx,
const std::string & timestamp, const std::vector<int> & prompt_tokens, const char * model_desc);
//
// KV cache utils
//
// Dump the KV cache view with the number of sequences per cell.
void dump_kv_cache_view(const llama_kv_cache_view & view, int row_size = 80);
// Dump the KV cache view showing individual sequences in each cell (long output).
void dump_kv_cache_view_seqs(const llama_kv_cache_view & view, int row_size = 40);

2
common/grammar-parser.cpp
View file

@ -190,7 +190,7 @@ namespace grammar_parser {
pos = parse_space(pos + 1, is_nested);
} else if (*pos == '*' || *pos == '+' || *pos == '?') { // repetition operator
if (last_sym_start == out_elements.size()) {
throw std::runtime_error(std::string("expecting preceeding item to */+/? at ") + pos);
throw std::runtime_error(std::string("expecting preceding item to */+/? at ") + pos);
}
// apply transformation to previous symbol (last_sym_start to end) according to

8
common/log.h
View file

@ -61,13 +61,13 @@
// #define LOG_TARGET stderr
// #include "log.h"
//
// The log target can also be redirected to a diffrent function
// The log target can also be redirected to a different function
// like so:
//
// #define LOG_TARGET log_handler_diffrent()
// #define LOG_TARGET log_handler_different()
// #include "log.h"
//
// FILE* log_handler_diffrent()
// FILE* log_handler_different()
// {
// return stderr;
// }
@ -421,7 +421,7 @@ inline FILE *log_handler2_impl(bool change = false, LogTriState append = LogTriS
// Disables logs entirely at runtime.
// Makes LOG() and LOG_TEE() produce no output,
// untill enabled back.
// until enabled back.
#define log_disable() log_disable_impl()
// INTERNAL, DO NOT USE

118
common/sampling.cpp
View file

@ -99,21 +99,67 @@ std::string llama_sampling_print(const llama_sampling_params & params) {
return std::string(result);
}
llama_token llama_sampling_sample(
std::string llama_sampling_order_print(const llama_sampling_params & params) {
std::string result = "CFG -> Penalties ";
if (params.mirostat == 0) {
for (auto s : params.samplers_sequence) {
switch (s) {
case 'k': result += "-> top_k "; break;
case 'f': result += "-> tfs_z "; break;
case 'y': result += "-> typical_p "; break;
case 'p': result += "-> top_p "; break;
case 'm': result += "-> min_p "; break;
case 't': result += "-> temp "; break;
default : break;
}
}
} else {
result += "-> mirostat ";
}
return result;
}
// no reasons to expose this function in header
static void sampler_queue(
struct llama_context * ctx_main,
const llama_sampling_params & params,
llama_token_data_array & cur_p,
size_t & min_keep) {
const int n_vocab = llama_n_vocab(llama_get_model(ctx_main));
const float temp = params.temp;
const int32_t top_k = params.top_k <= 0 ? n_vocab : params.top_k;
const float top_p = params.top_p;
const float min_p = params.min_p;
const float tfs_z = params.tfs_z;
const float typical_p = params.typical_p;
const std::string & samplers_sequence = params.samplers_sequence;
for (auto s : samplers_sequence) {
switch (s){
case 'k': llama_sample_top_k (ctx_main, &cur_p, top_k, min_keep); break;
case 'f': llama_sample_tail_free(ctx_main, &cur_p, tfs_z, min_keep); break;
case 'y': llama_sample_typical (ctx_main, &cur_p, typical_p, min_keep); break;
case 'p': llama_sample_top_p (ctx_main, &cur_p, top_p, min_keep); break;
case 'm': llama_sample_min_p (ctx_main, &cur_p, min_p, min_keep); break;
case 't': llama_sample_temp (ctx_main, &cur_p, temp); break;
default : break;
}
}
}
static llama_token llama_sampling_sample_impl(
struct llama_sampling_context * ctx_sampling,
struct llama_context * ctx_main,
struct llama_context * ctx_cfg,
const int idx) {
const int idx,
bool is_resampling) { // Add a parameter to indicate if we are resampling
const llama_sampling_params & params = ctx_sampling->params;
const int n_vocab = llama_n_vocab(llama_get_model(ctx_main));
const float temp = params.temp;
const int32_t top_k = params.top_k <= 0 ? n_vocab : params.top_k;
const float top_p = params.top_p;
const float min_p = params.min_p;
const float tfs_z = params.tfs_z;
const float typical_p = params.typical_p;
const int32_t penalty_last_n = params.penalty_last_n < 0 ? params.n_prev : params.penalty_last_n;
const float penalty_repeat = params.penalty_repeat;
const float penalty_freq = params.penalty_freq;
@ -128,8 +174,17 @@ llama_token llama_sampling_sample(
llama_token id = 0;
// Get a pointer to the logits
float * logits = llama_get_logits_ith(ctx_main, idx);
// Declare original_logits at the beginning of the function scope
std::vector<float> original_logits;
if (!is_resampling) {
// Only make a copy of the original logits if we are not in the resampling phase, not sure if I actually have to do this.
original_logits = std::vector<float>(logits, logits + llama_n_vocab(llama_get_model(ctx_main)));
}
// apply params.logit_bias map
for (auto it = params.logit_bias.begin(); it != params.logit_bias.end(); it++) {
logits[it->first] += it->second;
@ -148,12 +203,14 @@ llama_token llama_sampling_sample(
}
// apply penalties
if (!prev.empty()) {
const auto& penalty_tokens = params.use_penalty_prompt_tokens ? params.penalty_prompt_tokens : prev;
const int penalty_tokens_used_size = std::min((int)penalty_tokens.size(), penalty_last_n);
if (penalty_tokens_used_size) {
const float nl_logit = logits[llama_token_nl(llama_get_model(ctx_main))];
llama_sample_repetition_penalties(ctx_main, &cur_p,
prev.data() + prev.size() - penalty_last_n,
penalty_last_n, penalty_repeat, penalty_freq, penalty_present);
penalty_tokens.data() + penalty_tokens.size() - penalty_tokens_used_size,
penalty_tokens_used_size, penalty_repeat, penalty_freq, penalty_present);
if (!penalize_nl) {
for (size_t idx = 0; idx < cur_p.size; idx++) {
@ -165,7 +222,8 @@ llama_token llama_sampling_sample(
}
}
if (ctx_sampling->grammar != NULL) {
// If we are in the resampling phase, apply grammar checks before sampling logic
if (is_resampling && ctx_sampling->grammar != NULL) {
llama_sample_grammar(ctx_main, &cur_p, ctx_sampling->grammar);
}
@ -188,12 +246,7 @@ llama_token llama_sampling_sample(
// temperature sampling
size_t min_keep = std::max(1, params.n_probs);
llama_sample_top_k (ctx_main, &cur_p, top_k, min_keep);
llama_sample_tail_free(ctx_main, &cur_p, tfs_z, min_keep);
llama_sample_typical (ctx_main, &cur_p, typical_p, min_keep);
llama_sample_top_p (ctx_main, &cur_p, top_p, min_keep);
llama_sample_min_p (ctx_main, &cur_p, min_p, min_keep);
llama_sample_temp (ctx_main, &cur_p, temp);
sampler_queue(ctx_main, params, cur_p, min_keep);
id = llama_sample_token(ctx_main, &cur_p);
@ -212,9 +265,40 @@ llama_token llama_sampling_sample(
}
}
if (ctx_sampling->grammar != NULL && !is_resampling) {
// Create an array with a single token data element for the sampled id
llama_token_data single_token_data = {id, logits[id], 0.0f};
llama_token_data_array single_token_data_array = { &single_token_data, 1, false };
// Apply grammar constraints to the single token
llama_sample_grammar(ctx_main, &single_token_data_array, ctx_sampling->grammar);
// Check if the token is valid according to the grammar by seeing if its logit has been set to -INFINITY
bool is_valid = single_token_data_array.data[0].logit != -INFINITY;
// If the token is not valid according to the grammar, perform resampling
if (!is_valid) {
LOG("Resampling because token %d: '%s' does not meet grammar rules\n", id, llama_token_to_piece(ctx_main, id).c_str());
// Restore logits from the copy
std::copy(original_logits.begin(), original_logits.end(), logits);
return llama_sampling_sample_impl(ctx_sampling, ctx_main, ctx_cfg, idx, true); // Pass true for is_resampling
}
}
return id;
}
llama_token llama_sampling_sample(
struct llama_sampling_context * ctx_sampling,
struct llama_context * ctx_main,
struct llama_context * ctx_cfg,
const int idx) {
// Call the implementation function with is_resampling set to false by default
return llama_sampling_sample_impl(ctx_sampling, ctx_main, ctx_cfg, idx, false);
}
void llama_sampling_accept(
struct llama_sampling_context * ctx_sampling,
struct llama_context * ctx_main,

39
common/sampling.h
View file

@ -10,22 +10,23 @@
// sampling parameters
typedef struct llama_sampling_params {
int32_t n_prev = 64; // number of previous tokens to remember
int32_t n_probs = 0; // if greater than 0, output the probabilities of top n_probs tokens.
int32_t top_k = 40; // <= 0 to use vocab size
float top_p = 0.95f; // 1.0 = disabled
float min_p = 0.05f; // 0.0 = disabled
float tfs_z = 1.00f; // 1.0 = disabled
float typical_p = 1.00f; // 1.0 = disabled
float temp = 0.80f; // 1.0 = disabled
int32_t penalty_last_n = 64; // last n tokens to penalize (0 = disable penalty, -1 = context size)
float penalty_repeat = 1.10f; // 1.0 = disabled
float penalty_freq = 0.00f; // 0.0 = disabled
float penalty_present = 0.00f; // 0.0 = disabled
int32_t mirostat = 0; // 0 = disabled, 1 = mirostat, 2 = mirostat 2.0
float mirostat_tau = 5.00f; // target entropy
float mirostat_eta = 0.10f; // learning rate
bool penalize_nl = true; // consider newlines as a repeatable token
int32_t n_prev = 64; // number of previous tokens to remember
int32_t n_probs = 0; // if greater than 0, output the probabilities of top n_probs tokens.
int32_t top_k = 40; // <= 0 to use vocab size
float top_p = 0.95f; // 1.0 = disabled
float min_p = 0.05f; // 0.0 = disabled
float tfs_z = 1.00f; // 1.0 = disabled
float typical_p = 1.00f; // 1.0 = disabled
float temp = 0.80f; // 1.0 = disabled
int32_t penalty_last_n = 64; // last n tokens to penalize (0 = disable penalty, -1 = context size)
float penalty_repeat = 1.10f; // 1.0 = disabled
float penalty_freq = 0.00f; // 0.0 = disabled
float penalty_present = 0.00f; // 0.0 = disabled
int32_t mirostat = 0; // 0 = disabled, 1 = mirostat, 2 = mirostat 2.0
float mirostat_tau = 5.00f; // target entropy
float mirostat_eta = 0.10f; // learning rate
bool penalize_nl = true; // consider newlines as a repeatable token
std::string samplers_sequence = "kfypmt"; // top_k, tail_free, typical_p, top_p, min_p, temp
std::string grammar; // optional BNF-like grammar to constrain sampling
@ -35,6 +36,9 @@ typedef struct llama_sampling_params {
float cfg_scale = 1.f; // how strong is guidance
std::unordered_map<llama_token, float> logit_bias; // logit bias for specific tokens
std::vector<llama_token> penalty_prompt_tokens;
bool use_penalty_prompt_tokens = false;
} llama_sampling_params;
// general sampler context
@ -80,6 +84,9 @@ std::string llama_sampling_prev_str(llama_sampling_context * ctx_sampling, llama
// Print sampling parameters into a string
std::string llama_sampling_print(const llama_sampling_params & params);
// Print sampling order into a string
std::string llama_sampling_order_print(const llama_sampling_params & params);
// this is a common sampling function used across the examples for convenience
// it can serve as a starting point for implementing your own sampling function
// Note: When using multiple sequences, it is the caller's responsibility to call

18
common/train.cpp
View file

@ -71,7 +71,7 @@ void free_random_uniform_distribution(struct random_uniform_distribution * rnd)
struct ggml_tensor * randomize_tensor_normal(struct ggml_tensor * tensor, struct random_normal_distribution * rnd) {
float scale = 1.0f; // xavier
switch (tensor->n_dims) {
switch (ggml_n_dims(tensor)) {
case 1:
scale /= sqrtf((float) tensor->ne[0]);
for (int i0 = 0; i0 < tensor->ne[0]; i0++) {
@ -119,7 +119,7 @@ struct ggml_tensor * randomize_tensor_normal(struct ggml_tensor * tensor, struct
}
struct ggml_tensor * randomize_tensor_uniform(struct ggml_tensor * tensor, struct random_uniform_distribution * rnd) {
switch (tensor->n_dims) {
switch (ggml_n_dims(tensor)) {
case 1:
for (int i0 = 0; i0 < tensor->ne[0]; i0++) {
float * dst = (float *) ((char *) tensor->data + i0*tensor->nb[0]);
@ -183,25 +183,27 @@ float fclamp(const float v, const float min, const float max) {
}
void assert_shape_1d(struct ggml_tensor * tensor, int64_t ne0) {
GGML_ASSERT(tensor->n_dims == 1);
GGML_ASSERT(tensor->ne[0] == ne0);
GGML_ASSERT(tensor->ne[1] == 1);
GGML_ASSERT(tensor->ne[2] == 1);
GGML_ASSERT(tensor->ne[3] == 1);
}
void assert_shape_2d(struct ggml_tensor * tensor, int64_t ne0, int64_t ne1) {
GGML_ASSERT(tensor->n_dims == 2);
GGML_ASSERT(tensor->ne[0] == ne0);
GGML_ASSERT(tensor->ne[1] == ne1);
GGML_ASSERT(tensor->ne[2] == 1);
GGML_ASSERT(tensor->ne[3] == 1);
}
void assert_shape_3d(struct ggml_tensor * tensor, int64_t ne0, int64_t ne1, int64_t ne2) {
GGML_ASSERT(tensor->n_dims == 3);
GGML_ASSERT(tensor->ne[0] == ne0);
GGML_ASSERT(tensor->ne[1] == ne1);
GGML_ASSERT(tensor->ne[2] == ne2);
GGML_ASSERT(tensor->ne[3] == 1);
}
void assert_shape_4d(struct ggml_tensor * tensor, int64_t ne0, int64_t ne1, int64_t ne2, int64_t ne3) {
GGML_ASSERT(tensor->n_dims == 4);
GGML_ASSERT(tensor->ne[0] == ne0);
GGML_ASSERT(tensor->ne[1] == ne1);
GGML_ASSERT(tensor->ne[2] == ne2);
@ -225,8 +227,8 @@ int64_t get_example_targets_batch(
bool sample_random_offsets
) {
GGML_ASSERT(samples_count > 0);
GGML_ASSERT(tokens_input->n_dims == 2);
GGML_ASSERT(target_probs->n_dims == 3);
GGML_ASSERT(ggml_is_matrix(tokens_input));
GGML_ASSERT(ggml_is_3d(target_probs));
int64_t n_vocab = target_probs->ne[0];
int64_t n_tokens = tokens_input->ne[0];
int64_t n_batch = tokens_input->ne[1];

340
convert-hf-to-gguf.py
View file

@ -10,7 +10,7 @@ import re
import sys
from enum import IntEnum
from pathlib import Path
from typing import TYPE_CHECKING, Any, ContextManager, Iterator, cast
from typing import TYPE_CHECKING, Any, ContextManager, Iterator, cast, Optional
import numpy as np
import torch
@ -46,7 +46,7 @@ class Model:
self.part_names = self._get_part_names()
self.hparams = Model.load_hparams(self.dir_model)
self.model_arch = self._get_model_architecture()
self.gguf_writer = gguf.GGUFWriter(fname_out, gguf.MODEL_ARCH_NAMES[self.model_arch], endianess=self.endianess)
self.gguf_writer = gguf.GGUFWriter(fname_out, gguf.MODEL_ARCH_NAMES[self.model_arch], endianess=self.endianess, use_temp_file=False)
def set_vocab(self):
self._set_vocab_gpt2()
@ -59,7 +59,7 @@ class Model:
from safetensors import safe_open
ctx = cast(ContextManager[Any], safe_open(self.dir_model / part_name, framework="pt", device="cpu"))
else:
ctx = contextlib.nullcontext(torch.load(self.dir_model / part_name, map_location="cpu"))
ctx = contextlib.nullcontext(torch.load(str(self.dir_model / part_name), map_location="cpu", weights_only=True))
with ctx as model_part:
for name in model_part.keys():
@ -77,8 +77,18 @@ class Model:
self.gguf_writer.add_embedding_length(n_embd)
if (n_ff := self.hparams.get("intermediate_size")) is not None:
self.gguf_writer.add_feed_forward_length(n_ff)
if (n_head := self.hparams.get("num_attention_head")) is not None:
if (n_head := self.hparams.get("num_attention_heads")) is not None:
self.gguf_writer.add_head_count(n_head)
if (n_head_kv := self.hparams.get("num_key_value_heads")) is not None:
self.gguf_writer.add_head_count_kv(n_head_kv)
if (n_rms_eps := self.hparams.get("rms_norm_eps")) is not None:
self.gguf_writer.add_layer_norm_rms_eps(n_rms_eps)
if (n_experts := self.hparams.get("num_local_experts")) is not None:
self.gguf_writer.add_expert_count(n_experts)
if (n_experts_used := self.hparams.get("num_experts_per_tok")) is not None:
self.gguf_writer.add_expert_used_count(n_experts_used)
self.gguf_writer.add_parallel_residual(self.hparams.get("use_parallel_residual", True))
def write_tensors(self):
@ -170,6 +180,40 @@ class Model:
return DeepseekCoderModel
if model_architecture in ("StableLMEpochForCausalLM", "LlavaStableLMEpochForCausalLM"):
return StableLMModel
if model_architecture == "QWenLMHeadModel":
return QwenModel
if model_architecture == "MixtralForCausalLM":
return MixtralModel
if model_architecture == "PhiForCausalLM":
return Phi2Model
if model_architecture == "PlamoForCausalLM":
return PlamoModel
@staticmethod
def from_model_name(model_name: str):
model_name_lower = model_name.lower()
if model_name_lower in ("stablelmepoch", "llavastablelmepoch"):
return StableLMModel
if model_name_lower == "gptneox":
return GPTNeoXModel
if model_name_lower == "bloom":
return BloomModel
if model_name_lower == "mpt":
return MPTModel
if model_name_lower in ("baichuan"):
return BaichuanModel
if model_name_lower in ("falcon", "rw"):
return FalconModel
if model_name_lower == "gptbigcode":
return StarCoderModel
if model_name_lower == "gptrefact":
return RefactModel
if model_name_lower == "persimmon":
return PersimmonModel
if model_name_lower == "deepseekcoder":
return DeepseekCoderModel
if model_name_lower == "deepseekllm":
return DeepseekLLMModel
return Model
@staticmethod
@ -234,6 +278,14 @@ class Model:
return gguf.MODEL_ARCH.LLAMA
if arch in ("StableLMEpochForCausalLM", "LlavaStableLMEpochForCausalLM"):
return gguf.MODEL_ARCH.STABLELM
if arch == "QWenLMHeadModel":
return gguf.MODEL_ARCH.QWEN
if arch == "MixtralForCausalLM":
return gguf.MODEL_ARCH.LLAMA
if arch == "PhiForCausalLM":
return gguf.MODEL_ARCH.PHI2
if arch == "PlamoForCausalLM":
return gguf.MODEL_ARCH.PLAMO
raise NotImplementedError(f'Architecture "{arch}" not supported!')
@ -469,7 +521,11 @@ class MPTModel(Model):
data = data_torch.squeeze().numpy()
# map tensor names
new_name = tensor_map.get_name(name, try_suffixes=(".weight", ".bias"))
if "scales" in name:
new_name = tensor_map.get_name(name, try_suffixes=(".weight", ".bias", ".scales"))
new_name = new_name.replace("scales", "act.scales")
else:
new_name = tensor_map.get_name(name, try_suffixes=(".weight", ".bias"))
if new_name is None:
print(f"Can not map tensor {name!r}")
sys.exit()
@ -886,15 +942,246 @@ class StableLMModel(Model):
self.gguf_writer.add_parallel_residual(hparams["use_parallel_residual"] if "use_parallel_residual" in hparams else True)
self.gguf_writer.add_layer_norm_eps(1e-5)
class MixtralModel(Model):
def set_vocab(self):
self._set_vocab_sentencepiece()
class QwenModel(Model):
@staticmethod
def token_bytes_to_string(b):
from transformers.models.gpt2.tokenization_gpt2 import bytes_to_unicode
byte_encoder = bytes_to_unicode()
return ''.join([byte_encoder[ord(char)] for char in b.decode('latin-1')])
@staticmethod
def bpe(mergeable_ranks: dict[bytes, int], token: bytes, max_rank: Optional[int] = None) -> list[bytes]:
parts = [bytes([b]) for b in token]
while True:
min_idx = None
min_rank = None
for i, pair in enumerate(zip(parts[:-1], parts[1:])):
rank = mergeable_ranks.get(pair[0] + pair[1])
if rank is not None and (min_rank is None or rank < min_rank):
min_idx = i
min_rank = rank
if min_rank is None or (max_rank is not None and min_rank >= max_rank):
break
assert min_idx is not None
parts = parts[:min_idx] + [parts[min_idx] + parts[min_idx + 1]] + parts[min_idx + 2:]
return parts
def set_vocab(self):
dir_model = self.dir_model
hparams = self.hparams
tokens: list[bytearray] = []
toktypes: list[int] = []
from transformers import AutoTokenizer # type: ignore[attr-defined]
tokenizer = AutoTokenizer.from_pretrained(dir_model, trust_remote_code=True)
vocab_size = hparams["vocab_size"]
assert max(tokenizer.get_vocab().values()) < vocab_size
merges = []
vocab = {}
mergeable_ranks = tokenizer.mergeable_ranks
for token, rank in mergeable_ranks.items():
vocab[self.token_bytes_to_string(token)] = rank
if len(token) == 1:
continue
merged = QwenModel.bpe(mergeable_ranks, token, max_rank=rank)
assert len(merged) == 2
merges.append(' '.join(map(self.token_bytes_to_string, merged)))
reverse_vocab = {id_ : encoded_tok for encoded_tok, id_ in vocab.items()}
added_vocab = tokenizer.special_tokens
for i in range(vocab_size):
if i not in reverse_vocab:
pad_token = f"[PAD{i}]".encode("utf-8")
tokens.append(bytearray(pad_token))
toktypes.append(gguf.TokenType.USER_DEFINED)
elif reverse_vocab[i] in added_vocab:
tokens.append(reverse_vocab[i])
toktypes.append(gguf.TokenType.CONTROL)
else:
tokens.append(reverse_vocab[i])
toktypes.append(gguf.TokenType.NORMAL)
self.gguf_writer.add_tokenizer_model("gpt2")
self.gguf_writer.add_token_list(tokens)
self.gguf_writer.add_token_types(toktypes)
special_vocab = gguf.SpecialVocab(dir_model, load_merges=False)
special_vocab.merges = merges
special_vocab._set_special_token("bos", tokenizer.special_tokens["<|endoftext|>"])
special_vocab._set_special_token("eos", tokenizer.special_tokens["<|endoftext|>"])
special_vocab._set_special_token("unk", tokenizer.special_tokens["<|endoftext|>"])
special_vocab.add_to_gguf(self.gguf_writer)
def set_gguf_parameters(self):
self.gguf_writer.add_name("Qwen")
self.gguf_writer.add_context_length(self.hparams["max_position_embeddings"])
self.gguf_writer.add_block_count(self.hparams["num_hidden_layers"])
self.gguf_writer.add_embedding_length(self.hparams["hidden_size"])
self.gguf_writer.add_feed_forward_length(self.hparams["intermediate_size"])
self.gguf_writer.add_rope_freq_base(self.hparams["rotary_emb_base"])
self.gguf_writer.add_rope_dimension_count(self.hparams["hidden_size"] // self.hparams["num_attention_heads"])
self.gguf_writer.add_head_count(self.hparams["num_attention_heads"])
self.gguf_writer.add_layer_norm_rms_eps(self.hparams["layer_norm_epsilon"])
def write_tensors(self):
block_count = self.hparams["num_hidden_layers"]
model_kv = dict(self.get_tensors())
tensor_map = gguf.get_tensor_name_map(self.model_arch, block_count)
for name, data_torch in model_kv.items():
# we don't need these
if name.endswith(".rotary_emb.inv_freq"):
continue
old_dtype = data_torch.dtype
# convert any unsupported data types to float32
if data_torch.dtype not in (torch.float16, torch.float32):
data_torch = data_torch.to(torch.float32)
data = data_torch.squeeze().numpy()
# map tensor names
new_name = tensor_map.get_name(name, try_suffixes=(".weight", ".bias"))
if new_name is None:
print(f"Can not map tensor {name!r}")
sys.exit()
n_dims = len(data.shape)
data_dtype = data.dtype
# if f32 desired, convert any float16 to float32
if self.ftype == 0 and data_dtype == np.float16:
data = data.astype(np.float32)
# TODO: Why cant we use these float16 as-is? There should be not reason to store float16 as float32
if self.ftype == 1 and data_dtype == np.float16 and n_dims == 1:
data = data.astype(np.float32)
# if f16 desired, convert any float32 2-dim weight tensors to float16
if self.ftype == 1 and data_dtype == np.float32 and name.endswith(".weight") and n_dims == 2:
data = data.astype(np.float16)
print(f"{new_name}, n_dims = {n_dims}, {old_dtype} --> {data.dtype}")
self.gguf_writer.add_tensor(new_name, data)
class Phi2Model(Model):
def set_gguf_parameters(self):
block_count = self.hparams["n_layer"]
self.gguf_writer.add_name("Phi2")
self.gguf_writer.add_context_length(self.hparams["n_positions"])
self.gguf_writer.add_embedding_length(self.hparams["n_embd"])
self.gguf_writer.add_feed_forward_length(4 * self.hparams["n_embd"])
self.gguf_writer.add_block_count(block_count)
self.gguf_writer.add_head_count(self.hparams["n_head"])
self.gguf_writer.add_head_count_kv(self.hparams["n_head"])
self.gguf_writer.add_layer_norm_eps(self.hparams["layer_norm_epsilon"])
self.gguf_writer.add_rope_dimension_count(self.hparams["rotary_dim"])
self.gguf_writer.add_file_type(self.ftype)
self.gguf_writer.add_add_bos_token(False)
class PlamoModel(Model):
def set_vocab(self):
self._set_vocab_sentencepiece()
def set_gguf_parameters(self):
hparams = self.hparams
block_count = hparams["num_hidden_layers"]
self.gguf_writer.add_name("PLaMo")
self.gguf_writer.add_context_length(4096) # not in config.json
self.gguf_writer.add_embedding_length(hparams["hidden_size"])
self.gguf_writer.add_feed_forward_length(hparams["intermediate_size"])
self.gguf_writer.add_block_count(block_count)
self.gguf_writer.add_head_count(hparams["num_attention_heads"])
self.gguf_writer.add_head_count_kv(5) # hparams["num_key_value_heads"]) is wrong
self.gguf_writer.add_layer_norm_rms_eps(hparams["rms_norm_eps"])
def shuffle_attn_q_weight(self, data_torch):
assert data_torch.size() == (5120, 5120)
data_torch = data_torch.reshape(8, 5, 128, 5120)
data_torch = torch.permute(data_torch, (1, 0, 2, 3))
data_torch = torch.reshape(data_torch, (5120, 5120))
return data_torch
def shuffle_attn_output_weight(self, data_torch):
assert data_torch.size() == (5120, 5120)
data_torch = data_torch.reshape(5120, 8, 5, 128)
data_torch = torch.permute(data_torch, (0, 2, 1, 3))
data_torch = torch.reshape(data_torch, (5120, 5120))
return data_torch
def write_tensors(self):
block_count = self.hparams.get("num_layers", self.hparams.get("num_hidden_layers"))
tensor_map = gguf.get_tensor_name_map(self.model_arch, block_count)
for name, data_torch in self.get_tensors():
if "self_attn.rotary_emb.inv_freq" in name:
continue
# map tensor names
new_name = tensor_map.get_name(name, try_suffixes=(".weight", ".bias"))
if new_name is None:
print(f"Can not map tensor {name!r}")
sys.exit()
# shuffle for broadcasting of gqa in ggml_mul_mat
if new_name.endswith("attn_q.weight"):
data_torch = self.shuffle_attn_q_weight(data_torch)
elif new_name.endswith("attn_output.weight"):
data_torch = self.shuffle_attn_output_weight(data_torch)
old_dtype = data_torch.dtype
# convert any unsupported data types to float32
if data_torch.dtype not in (torch.float16, torch.float32):
data_torch = data_torch.to(torch.float32)
data = data_torch.squeeze().numpy()
n_dims = len(data.shape)
data_dtype = data.dtype
# if f32 desired, convert any float16 to float32
if self.ftype == 0 and data_dtype == np.float16:
data = data.astype(np.float32)
# TODO: Why cant we use these float16 as-is? There should be not reason to store float16 as float32
if self.ftype == 1 and data_dtype == np.float16 and n_dims == 1:
data = data.astype(np.float32)
# if f16 desired, convert any float32 2-dim weight tensors to float16
if self.ftype == 1 and data_dtype == np.float32 and name.endswith(".weight") and n_dims == 2:
data = data.astype(np.float16)
print(f"{new_name}, n_dims = {n_dims}, {old_dtype} --> {data.dtype}")
self.gguf_writer.add_tensor(new_name, data)
###### CONVERSION LOGIC ######
def parse_args() -> argparse.Namespace:
parser = argparse.ArgumentParser(description="Convert a huggingface model to a GGML compatible file")
parser = argparse.ArgumentParser(
description="Convert a huggingface model to a GGML compatible file")
parser.add_argument(
"--vocab-only", action="store_true",
help="extract only the vocab",
)
parser.add_argument(
"--awq-path", type=Path, default=None,
help="Path to scale awq cache file")
parser.add_argument(
"--outfile", type=Path,
help="path to write to; default: based on input",
@ -916,6 +1203,20 @@ def parse_args() -> argparse.Namespace:
args = parse_args()
dir_model = args.model
if args.awq_path:
sys.path.insert(1, str(Path(__file__).parent / 'awq-py'))
from awq.apply_awq import add_scale_weights
tmp_model_path = args.model / "weighted_model"
dir_model = tmp_model_path
if tmp_model_path.is_dir():
print(f"{tmp_model_path} exists as a weighted model.")
else:
tmp_model_path.mkdir(parents=True, exist_ok=True)
print("Saving new weighted model ...")
add_scale_weights(str(args.model), str(args.awq_path), str(tmp_model_path))
print(f"Saved weighted model at {tmp_model_path}.")
if not dir_model.is_dir():
print(f'Error: {args.model} is not a directory', file=sys.stderr)
sys.exit(1)
@ -935,20 +1236,21 @@ print(f"Loading model: {dir_model.name}")
hparams = Model.load_hparams(dir_model)
model_class = Model.from_model_name(args.model_name) if args.model_name else Model.from_model_architecture(hparams["architectures"][0])
model_instance = model_class(dir_model, ftype_map[args.outtype], fname_out, args.bigendian)
with torch.inference_mode():
model_class = Model.from_model_name(args.model_name) if args.model_name else Model.from_model_architecture(hparams["architectures"][0])
model_instance = model_class(dir_model, ftype_map[args.outtype], fname_out, args.bigendian)
print("Set model parameters")
model_instance.set_gguf_parameters()
print("Set model parameters")
model_instance.set_gguf_parameters()
print("Set model tokenizer")
model_instance.set_vocab()
print("Set model tokenizer")
model_instance.set_vocab()
if args.vocab_only:
print(f"Exporting model vocab to '{fname_out}'")
model_instance.write_vocab()
else:
print(f"Exporting model to '{fname_out}'")
model_instance.write()
if args.vocab_only:
print(f"Exporting model vocab to '{fname_out}'")
model_instance.write_vocab()
else:
print(f"Exporting model to '{fname_out}'")
model_instance.write()
print(f"Model successfully exported to '{fname_out}'")
print(f"Model successfully exported to '{fname_out}'")

86
convert-lora-to-ggml.py
View file

@ -3,7 +3,6 @@ from __future__ import annotations
import json
import os
import re
import struct
import sys
from typing import Any, BinaryIO, Sequence
@ -11,43 +10,15 @@ from typing import Any, BinaryIO, Sequence
import numpy as np
import torch
from pathlib import Path
if 'NO_LOCAL_GGUF' not in os.environ:
sys.path.insert(1, str(Path(__file__).parent / 'gguf-py' / 'gguf'))
import gguf
NUMPY_TYPE_TO_FTYPE: dict[str, int] = {"float32": 0, "float16": 1}
HF_SUBLAYER_TO_GGML = {
"self_attn.q_proj": "attn_q",
"self_attn.k_proj": "attn_k",
"self_attn.v_proj": "attn_v",
"self_attn.o_proj": "attn_output",
"mlp.gate_proj": "ffn_gate",
"mlp.down_proj": "ffn_down",
"mlp.up_proj": "ffn_up",
"input_layernorm": "attn_norm",
"post_attention_layernorm": "ffn_norm",
}
def translate_tensor_name(t: str) -> str:
match = re.match(r".*layers\.(\d+)\.(\w+\.\w+)\.lora_(A|B)\.weight", t)
if match:
nn = match.group(1)
sub_layer = match.group(2)
lora_type = match.group(3)
sub_layer_renamed = HF_SUBLAYER_TO_GGML.get(sub_layer)
if sub_layer_renamed is None:
print(f"Error: unrecognized sub-layer {sub_layer} in tensor {t}")
sys.exit(1)
output_string = (
f"blk.{nn}.{HF_SUBLAYER_TO_GGML[sub_layer]}.weight.lora{lora_type}"
)
return output_string
else:
print(f"Error: unrecognized tensor {t}")
sys.exit(1)
def write_file_header(fout: BinaryIO, params: dict[str, Any]) -> None:
fout.write(b"ggla"[::-1]) # magic (ggml lora)
fout.write(struct.pack("i", 1)) # file version
@ -61,9 +32,7 @@ def write_file_header(fout: BinaryIO, params: dict[str, Any]) -> None:
fout.write(struct.pack("i", int(params["lora_alpha"])))
def write_tensor_header(
self, name: str, shape: Sequence[int], data_type: np.dtype[Any]
) -> None:
def write_tensor_header(fout: BinaryIO, name: str, shape: Sequence[int], data_type: np.dtype[Any]) -> None:
sname = name.encode("utf-8")
fout.write(
struct.pack(
@ -78,11 +47,12 @@ def write_tensor_header(
fout.seek((fout.tell() + 31) & -32)
if len(sys.argv) != 2:
print(f"Usage: python {sys.argv[0]} <path>")
if len(sys.argv) < 2:
print(f"Usage: python {sys.argv[0]} <path> [arch]")
print(
"Path must contain HuggingFace PEFT LoRA files 'adapter_config.json' and 'adapter_model.bin'"
)
print(f"Arch must be one of {list(gguf.MODEL_ARCH_NAMES.values())} (default: llama)")
sys.exit(1)
input_json = os.path.join(sys.argv[1], "adapter_config.json")
@ -90,6 +60,14 @@ input_model = os.path.join(sys.argv[1], "adapter_model.bin")
output_path = os.path.join(sys.argv[1], "ggml-adapter-model.bin")
model = torch.load(input_model, map_location="cpu")
arch_name = sys.argv[2] if len(sys.argv) == 3 else "llama"
if arch_name not in gguf.MODEL_ARCH_NAMES.values():
print(f"Error: unsupported architecture {arch_name}")
sys.exit(1)
arch = list(gguf.MODEL_ARCH_NAMES.keys())[list(gguf.MODEL_ARCH_NAMES.values()).index(arch_name)]
name_map = gguf.TensorNameMap(arch, 200) # 200 layers ought to be enough for anyone
with open(input_json, "r") as f:
params = json.load(f)
@ -117,6 +95,7 @@ with open(output_path, "wb") as fout:
write_file_header(fout, params)
for k, v in model.items():
orig_k = k
if k.endswith(".default.weight"):
k = k.replace(".default.weight", ".weight")
if k in ["llama_proj.weight", "llama_proj.bias"]:
@ -129,7 +108,32 @@ with open(output_path, "wb") as fout:
v = v.float()
t = v.detach().numpy()
tname = translate_tensor_name(k)
prefix = "base_model.model."
if k.startswith(prefix):
k = k[len(prefix) :]
lora_suffixes = (".lora_A.weight", ".lora_B.weight")
if k.endswith(lora_suffixes):
suffix = k[-len(lora_suffixes[0]):]
k = k[: -len(lora_suffixes[0])]
else:
print(f"Error: unrecognized tensor name {orig_k}")
sys.exit(1)
tname = name_map.get_name(k)
if tname is None:
print(f"Error: could not map tensor name {orig_k}")
print(" Note: the arch parameter must be specified if the model is not llama")
sys.exit(1)
if suffix == ".lora_A.weight":
tname += ".weight.loraA"
elif suffix == ".lora_B.weight":
tname += ".weight.loraB"
else:
assert False
print(f"{k} => {tname} {t.shape} {t.dtype} {t.nbytes/1024/1024:.2f}MB")
write_tensor_header(fout, tname, t.shape, t.dtype)
t.tofile(fout)

417
convert.py Normal file → Executable file
View file

@ -10,6 +10,7 @@ import itertools
import json
import math
import mmap
import os
import pickle
import re
import signal
@ -18,15 +19,15 @@ import sys
import time
import zipfile
from abc import ABCMeta, abstractmethod
from collections import OrderedDict
from concurrent.futures import ProcessPoolExecutor, ThreadPoolExecutor
from dataclasses import dataclass
from pathlib import Path
from typing import IO, TYPE_CHECKING, Any, Callable, Iterable, Literal, TypeVar
from typing import IO, TYPE_CHECKING, Any, Callable, Iterable, Literal, Optional, TypeVar, cast
import numpy as np
from sentencepiece import SentencePieceProcessor
import os
if 'NO_LOCAL_GGUF' not in os.environ:
sys.path.insert(1, str(Path(__file__).parent / 'gguf-py'))
import gguf
@ -42,6 +43,7 @@ NDArray: TypeAlias = 'np.ndarray[Any, Any]'
ARCH = gguf.MODEL_ARCH.LLAMA
DEFAULT_CONCURRENCY = 8
#
# data types
#
@ -62,10 +64,10 @@ class UnquantizedDataType(DataType):
pass
DT_F16 = UnquantizedDataType('F16', dtype = np.dtype(np.float16), valid_conversions = ['F32', 'Q8_0'])
DT_F32 = UnquantizedDataType('F32', dtype = np.dtype(np.float32), valid_conversions = ['F16', 'Q8_0'])
DT_I32 = UnquantizedDataType('I32', dtype = np.dtype(np.int16), valid_conversions = [])
DT_BF16 = UnquantizedDataType('BF16', dtype = np.dtype(np.uint16), valid_conversions = ['F32', 'F16', 'Q8_0'])
DT_F16 = UnquantizedDataType('F16', dtype = np.dtype(np.float16), valid_conversions = ['F32', 'Q8_0'])
DT_F32 = UnquantizedDataType('F32', dtype = np.dtype(np.float32), valid_conversions = ['F16', 'Q8_0'])
DT_I32 = UnquantizedDataType('I32', dtype = np.dtype(np.int16), valid_conversions = [])
DT_BF16 = UnquantizedDataType('BF16', dtype = np.dtype(np.uint16), valid_conversions = ['F32', 'F16', 'Q8_0'])
@dataclass(frozen=True)
@ -151,14 +153,16 @@ GGML_FILE_TYPE_TO_DATA_TYPE: dict[GGMLFileType, DataType] = {
@dataclass
class Params:
n_vocab: int
n_embd: int
n_layer: int
n_ctx: int
n_ff: int
n_head: int
n_head_kv: int
f_norm_eps: float
n_vocab: int
n_embd: int
n_layer: int
n_ctx: int
n_ff: int
n_head: int
n_head_kv: int
n_experts: int | None = None
n_experts_used: int | None = None
f_norm_eps: float | None = None
rope_scaling_type: gguf.RopeScalingType | None = None
f_rope_freq_base: float | None = None
@ -233,6 +237,13 @@ class Params:
raise Exception("failed to guess 'n_ctx'. This model is unknown or unsupported.\n"
"Suggestion: provide 'config.json' of the model in the same directory containing model files.")
n_experts = None
n_experts_used = None
if "num_local_experts" in config:
n_experts = config["num_local_experts"]
n_experts_used = config["num_experts_per_tok"]
return Params(
n_vocab = config["vocab_size"],
n_embd = config["hidden_size"],
@ -241,6 +252,8 @@ class Params:
n_ff = config["intermediate_size"],
n_head = (n_head := config["num_attention_heads"]),
n_head_kv = config.get("num_key_value_heads", n_head),
n_experts = n_experts,
n_experts_used = n_experts_used,
f_norm_eps = config["rms_norm_eps"],
f_rope_freq_base = config.get("rope_theta"),
rope_scaling_type = rope_scaling_type,
@ -255,8 +268,15 @@ class Params:
def loadOriginalParamsJson(model: LazyModel, config_path: Path) -> Params:
config = json.load(open(config_path))
n_experts = None
n_experts_used = None
f_rope_freq_base = None
# hack to determine LLaMA v1 vs v2 vs CodeLlama
if config.get("rope_theta") == 1000000:
if config.get("moe"):
# Mixtral
n_ctx = 32768
elif config.get("rope_theta") == 1000000:
# CodeLlama
n_ctx = 16384
elif config["norm_eps"] == 1e-05:
@ -266,16 +286,27 @@ class Params:
# LLaMA v1
n_ctx = 2048
if "layers.0.feed_forward.w1.weight" in model:
n_ff = model["layers.0.feed_forward.w1.weight"].shape[0]
if config.get("moe"):
n_ff = model["layers.0.feed_forward.experts.0.w1.weight"].shape[0]
n_experts = config["moe"]["num_experts"]
n_experts_used = config["moe"]["num_experts_per_tok"]
f_rope_freq_base = 1e6
return Params(
n_vocab = config.get("vocab_size", model["tok_embeddings.weight"].shape[0]),
n_vocab = model["tok_embeddings.weight"].shape[0],
n_embd = config["dim"],
n_layer = config["n_layers"],
n_ctx = n_ctx,
n_ff = model["layers.0.feed_forward.w1.weight"].shape[0],
n_ff = n_ff,
n_head = (n_head := config["n_heads"]),
n_head_kv = config.get("n_kv_heads", n_head),
n_experts = n_experts,
n_experts_used = n_experts_used,
f_norm_eps = config["norm_eps"],
f_rope_freq_base = config.get("rope_theta"),
f_rope_freq_base = config.get("rope_theta", f_rope_freq_base),
)
@staticmethod
@ -297,127 +328,140 @@ class Params:
return params
#
# vocab
#
class VocabLoader:
def __init__(self, params: Params, fname_tokenizer: Path) -> None:
try:
from transformers import AutoTokenizer
except ImportError as e:
raise ImportError(
"To use VocabLoader, please install the `transformers` package. "
"You can install it with `pip install transformers`."
) from e
class BpeVocab:
def __init__(self, fname_tokenizer: Path, fname_added_tokens: Path | None) -> None:
self.bpe_tokenizer = json.loads(open(str(fname_tokenizer), encoding="utf-8").read())
added_tokens: dict[str, int]
if fname_added_tokens is not None:
# FIXME: Verify that added tokens here _cannot_ overlap with the main vocab.
added_tokens = json.load(open(fname_added_tokens, encoding="utf-8"))
try:
self.tokenizer = AutoTokenizer.from_pretrained(str(fname_tokenizer), trust_remote_code=True)
except ValueError:
self.tokenizer = AutoTokenizer.from_pretrained(str(fname_tokenizer), use_fast=False, trust_remote_code=True)
self.added_tokens_dict: OrderedDict[str, int] = OrderedDict()
for tok, tokidx in sorted(self.tokenizer.get_added_vocab().items(), key=lambda x: x[1]):
if tokidx >= params.n_vocab or tokidx < self.tokenizer.vocab_size:
continue
self.added_tokens_dict[tok] = tokidx
self.unk_token_id: int = self.tokenizer.unk_token_id
self.specials: dict[str, int] = {
tok: self.tokenizer.get_vocab()[tok]
for tok in self.tokenizer.all_special_tokens
}
self.special_ids: set[int] = set(self.tokenizer.all_special_ids)
self.reverse_vocab = {id: encoded_tok for encoded_tok, id in self.tokenizer.get_vocab().items()}
self.vocab_size_base: int = self.tokenizer.vocab_size
self.vocab_size: int = self.vocab_size_base + len(self.added_tokens_dict)
self.fname_tokenizer: Path = fname_tokenizer
vocab_file = "tokenizer.model"
path_candidate = find_vocab_file_path(self.fname_tokenizer, vocab_file)
if path_candidate is not None:
self.spm = SentencePieceProcessor(str(path_candidate))
print(self.spm.vocab_size(), self.vocab_size_base)
else:
# Fall back to trying to find the added tokens in tokenizer.json
tokenizer_json_file = fname_tokenizer.parent / 'tokenizer.json'
if not tokenizer_json_file.is_file():
added_tokens = {}
else:
tokenizer_json = json.load(open(tokenizer_json_file, encoding="utf-8"))
added_tokens = dict(
(item['content'], item['id'])
for item in tokenizer_json.get('added_tokens', [])
# Added tokens here can be duplicates of the main vocabulary.
if item['content'] not in self.bpe_tokenizer)
self.spm = None
vocab_size: int = len(self.bpe_tokenizer)
expected_ids = list(range(vocab_size, vocab_size + len(added_tokens)))
actual_ids = sorted(added_tokens.values())
if expected_ids != actual_ids:
expected_end_id = vocab_size + len(actual_ids) - 1
raise Exception(f"Expected the {len(actual_ids)} added token ID(s) to be sequential in the range {vocab_size} - {expected_end_id}; got {actual_ids}")
def hf_tokens(self) -> Iterable[tuple[bytes, float, gguf.TokenType]]:
added_tokens_ids = set(self.added_tokens_dict.values())
items = sorted(added_tokens.items(), key=lambda text_idx: text_idx[1])
self.added_tokens_list = [text for (text, idx) in items]
self.vocab_size_base: int = vocab_size
self.vocab_size: int = self.vocab_size_base + len(self.added_tokens_list)
self.fname_tokenizer = fname_tokenizer
self.fname_added_tokens = fname_added_tokens
for i in range(self.vocab_size_base):
if i in added_tokens_ids:
continue
def bpe_tokens(self) -> Iterable[tuple[bytes, float, gguf.TokenType]]:
tokenizer = self.bpe_tokenizer
reverse_vocab = {id: encoded_tok for encoded_tok, id in tokenizer.items()}
text = self.reverse_vocab[i].encode("utf-8")
yield text, self.get_token_score(i), self.get_token_type(i)
for i, _ in enumerate(tokenizer):
yield reverse_vocab[i], 0.0, gguf.TokenType.NORMAL
def get_token_type(self, token_id: int) -> gguf.TokenType:
toktype = gguf.TokenType.NORMAL
def added_tokens(self) -> Iterable[tuple[bytes, float, gguf.TokenType]]:
for text in self.added_tokens_list:
score = -1000.0
yield text.encode("utf-8"), score, gguf.TokenType.CONTROL
def all_tokens(self) -> Iterable[tuple[bytes, float, gguf.TokenType]]:
yield from self.bpe_tokens()
yield from self.added_tokens()
def __repr__(self) -> str:
return f"<BpeVocab with {self.vocab_size_base} base tokens and {len(self.added_tokens_list)} added tokens>"
class SentencePieceVocab:
def __init__(self, fname_tokenizer: Path, fname_added_tokens: Path | None) -> None:
self.sentencepiece_tokenizer = SentencePieceProcessor(str(fname_tokenizer))
added_tokens: dict[str, int]
if fname_added_tokens is not None:
added_tokens = json.load(open(fname_added_tokens, encoding="utf-8"))
else:
added_tokens = {}
vocab_size: int = self.sentencepiece_tokenizer.vocab_size()
new_tokens = {id: piece for piece, id in added_tokens.items() if id >= vocab_size}
expected_new_ids = list(range(vocab_size, vocab_size + len(new_tokens)))
actual_new_ids = sorted(new_tokens.keys())
if expected_new_ids != actual_new_ids:
raise ValueError(f"Expected new token IDs {expected_new_ids} to be sequential; got {actual_new_ids}")
# Token pieces that were added to the base vocabulary.
self.added_tokens_list = [new_tokens[id] for id in actual_new_ids]
self.vocab_size_base = vocab_size
self.vocab_size = self.vocab_size_base + len(self.added_tokens_list)
self.fname_tokenizer = fname_tokenizer
self.fname_added_tokens = fname_added_tokens
def sentencepiece_tokens(self) -> Iterable[tuple[bytes, float, gguf.TokenType]]:
tokenizer = self.sentencepiece_tokenizer
for i in range(tokenizer.vocab_size()):
piece = tokenizer.id_to_piece(i)
text: bytes = piece.encode("utf-8")
score: float = tokenizer.get_score(i)
toktype = gguf.TokenType.NORMAL
if tokenizer.is_unknown(i):
if self.spm is not None and token_id < self.spm.vocab_size():
if self.spm.is_unknown(token_id):
toktype = gguf.TokenType.UNKNOWN
if tokenizer.is_control(i):
if self.spm.is_control(token_id):
toktype = gguf.TokenType.CONTROL
# NOTE: I think added_tokens are user defined.
# ref: https://github.com/google/sentencepiece/blob/master/src/sentencepiece_model.proto
# if tokenizer.is_user_defined(i): toktype = gguf.TokenType.USER_DEFINED
if tokenizer.is_unused(i):
if self.spm.is_unused(token_id):
toktype = gguf.TokenType.UNUSED
if tokenizer.is_byte(i):
if self.spm.is_byte(token_id):
toktype = gguf.TokenType.BYTE
else:
token = self.reverse_vocab[token_id]
if token_id == self.unk_token_id:
toktype = gguf.TokenType.UNKNOWN
elif token_id in self.special_ids:
toktype = gguf.TokenType.CONTROL
elif len(token) == 6 and token.startswith("<0x") and token.endswith(">"):
toktype = gguf.TokenType.BYTE
yield text, score, toktype
return toktype
def get_token_score(self, token_id: int) -> float:
if self.spm is not None and token_id < self.spm.vocab_size():
return cast(float, self.spm.get_score(token_id))
return 0.0
def added_tokens(self) -> Iterable[tuple[bytes, float, gguf.TokenType]]:
for text in self.added_tokens_list:
score = -1000.0
yield text.encode("utf-8"), score, gguf.TokenType.USER_DEFINED
for text in self.added_tokens_dict:
if text in self.specials:
toktype = self.get_token_type(self.specials[text])
score = self.get_token_score(self.specials[text])
else:
toktype = gguf.TokenType.USER_DEFINED
score = -1000.0
yield text.encode("utf-8"), score, toktype
def has_newline_token(self) -> bool:
return '<0x0A>' in self.tokenizer.vocab or '\n' in self.tokenizer.vocab
def all_tokens(self) -> Iterable[tuple[bytes, float, gguf.TokenType]]:
yield from self.sentencepiece_tokens()
yield from self.hf_tokens()
yield from self.added_tokens()
def get_vocab_type(self) -> str:
path_candidates = []
vocab_file = "tokenizer.model"
path_candidates.append(vocab_file)
path_candidate = find_vocab_file_path(self.fname_tokenizer, vocab_file)
if path_candidate is not None:
return "llama"
vocab_file = "vocab.json"
path_candidates.append(vocab_file)
path_candidate = find_vocab_file_path(self.fname_tokenizer, vocab_file)
if path_candidate is not None:
return "gpt2"
vocab_file = "tokenizer.json"
path_candidates.append(vocab_file)
path_candidate = find_vocab_file_path(self.fname_tokenizer, vocab_file)
if path_candidate:
if not self.has_newline_token():
return "gpt2"
return "llama"
raise FileNotFoundError(
f"Could not find {path_candidates} in {self.fname_tokenizer} or its parent; "
"if it's in another directory, pass the directory as --vocab-dir"
)
def __repr__(self) -> str:
return f"<SentencePieceVocab with {self.vocab_size_base} base tokens and {len(self.added_tokens_list)} added tokens>"
return f"<VocabLoader with {self.vocab_size_base} base tokens and {len(self.added_tokens_dict)} added tokens>"
Vocab: TypeAlias = 'BpeVocab | SentencePieceVocab'
Vocab: TypeAlias = 'VocabLoader'
#
# data loading
@ -585,7 +629,7 @@ def merge_multifile_models(models_plus: list[ModelPlus]) -> ModelPlus:
if any("model.embed_tokens.weight" in mp.model for mp in models_plus):
# Transformers models put different tensors in different files, but
# don't split indivdual tensors between files.
# don't split individual tensors between files.
model: LazyModel = {}
for mp in models_plus:
model.update(mp.model)
@ -678,7 +722,7 @@ class LazyUnpickler(pickle.Unpickler):
return func(*args)
CLASSES: dict[tuple[str, str], Any] = {
# getattr used here as a workaround for mypy not being smart enough to detrmine
# getattr used here as a workaround for mypy not being smart enough to determine
# the staticmethods have a __func__ attribute.
('torch._tensor', '_rebuild_from_type_v2'): getattr(rebuild_from_type_v2, '__func__'),
('torch._utils', '_rebuild_tensor_v2'): getattr(lazy_rebuild_tensor_v2, '__func__'),
@ -794,20 +838,27 @@ def bounded_parallel_map(func: Callable[[In], Out], iterable: Iterable[In], conc
yield result
def check_vocab_size(params: Params, vocab: Vocab) -> None:
def check_vocab_size(params: Params, vocab: Vocab, pad_vocab: bool = False) -> None:
if params.n_vocab != vocab.vocab_size:
assert isinstance(vocab, BpeVocab) or isinstance(vocab, SentencePieceVocab)
if params.n_vocab == vocab.vocab_size_base:
if params.n_vocab == vocab.vocab_size:
print("Ignoring added_tokens.json since model matches vocab size without it.")
vocab.added_tokens_list = []
vocab.vocab_size = vocab.vocab_size_base
vocab.added_tokens_dict = OrderedDict()
vocab.vocab_size = vocab.vocab_size
return
if pad_vocab and params.n_vocab > vocab.vocab_size:
pad_count = params.n_vocab - vocab.vocab_size
print(f'Padding vocab with {pad_count} token(s) - <dummy00001> through <dummy{pad_count:05}>')
for i in range(1, (params.n_vocab - vocab.vocab_size) + 1):
vocab.added_tokens_dict[f'<dummy{i:05}>'] = -1
vocab.vocab_size = params.n_vocab
return
msg = f"Vocab size mismatch (model has {params.n_vocab}, but {vocab.fname_tokenizer}"
if vocab.fname_added_tokens is not None:
msg += f" combined with {vocab.fname_added_tokens}"
msg += f" has {vocab.vocab_size})."
if vocab.vocab_size < params.n_vocab < vocab.vocab_size + 20 and vocab.fname_added_tokens is None:
if vocab.vocab_size < params.n_vocab < vocab.vocab_size + 20:
msg += f" Most likely you are missing added_tokens.json (should be in {vocab.fname_tokenizer.parent})."
if vocab.vocab_size < params.n_vocab:
msg += " Possibly try using the --padvocab option."
raise Exception(msg)
@ -832,7 +883,17 @@ class OutputFile:
self.gguf.add_rope_dimension_count(params.n_embd // params.n_head)
self.gguf.add_head_count (params.n_head)
self.gguf.add_head_count_kv (params.n_head_kv)
self.gguf.add_layer_norm_rms_eps (params.f_norm_eps)
if params.n_experts:
self.gguf.add_expert_count(params.n_experts)
if params.n_experts_used:
self.gguf.add_expert_used_count(params.n_experts_used)
if params.f_norm_eps:
self.gguf.add_layer_norm_rms_eps(params.f_norm_eps)
else:
raise ValueError('f_norm_eps is None')
if params.f_rope_freq_base is not None:
self.gguf.add_rope_freq_base(params.f_rope_freq_base)
@ -861,12 +922,8 @@ class OutputFile:
scores.append(score)
toktypes.append(toktype)
if isinstance(vocab, SentencePieceVocab):
self.gguf.add_tokenizer_model("llama")
elif isinstance(vocab, BpeVocab):
self.gguf.add_tokenizer_model("gpt2")
else:
raise ValueError('Unknown vocab type: Not BpeVocab or SentencePieceVocab')
vocab_type = vocab.get_vocab_type()
self.gguf.add_tokenizer_model(vocab_type)
self.gguf.add_token_list(tokens)
self.gguf.add_token_scores(scores)
self.gguf.add_token_types(toktypes)
@ -892,8 +949,12 @@ class OutputFile:
self.gguf.close()
@staticmethod
def write_vocab_only(fname_out: Path, params: Params, vocab: Vocab, svocab: gguf.SpecialVocab, endianess:gguf.GGUFEndian = gguf.GGUFEndian.LITTLE) -> None:
check_vocab_size(params, vocab)
def write_vocab_only(
fname_out: Path, params: Params, vocab: Vocab, svocab: gguf.SpecialVocab,
endianess: gguf.GGUFEndian = gguf.GGUFEndian.LITTLE,
pad_vocab: bool = False,
) -> None:
check_vocab_size(params, vocab, pad_vocab = pad_vocab)
of = OutputFile(fname_out, endianess=endianess)
@ -920,8 +981,13 @@ class OutputFile:
return dt.quantize(arr)
@staticmethod
def write_all(fname_out: Path, ftype: GGMLFileType, params: Params, model: LazyModel, vocab: Vocab, svocab: gguf.SpecialVocab, concurrency: int = DEFAULT_CONCURRENCY, endianess: gguf.GGUFEndian = gguf.GGUFEndian.LITTLE) -> None:
check_vocab_size(params, vocab)
def write_all(
fname_out: Path, ftype: GGMLFileType, params: Params, model: LazyModel, vocab: Vocab, svocab: gguf.SpecialVocab,
concurrency: int = DEFAULT_CONCURRENCY,
endianess: gguf.GGUFEndian = gguf.GGUFEndian.LITTLE,
pad_vocab: bool = False,
) -> None:
check_vocab_size(params, vocab, pad_vocab = pad_vocab)
of = OutputFile(fname_out, endianess=endianess)
@ -956,7 +1022,7 @@ class OutputFile:
def pick_output_type(model: LazyModel, output_type_str: str | None) -> GGMLFileType:
wq_type = model[gguf.TENSOR_NAMES[gguf.MODEL_TENSOR.ATTN_Q].format(bid=0) +".weight"].data_type
wq_type = model[gguf.TENSOR_NAMES[gguf.MODEL_TENSOR.ATTN_Q].format(bid=0) + ".weight"].data_type
if output_type_str == "f32" or (output_type_str is None and wq_type == DT_F32):
return GGMLFileType.AllF32
@ -1079,35 +1145,17 @@ def load_some_model(path: Path) -> ModelPlus:
return model_plus
def load_vocab(path: Path, vocabtype: str | None) -> Vocab:
# Be extra-friendly and accept either a file or a directory. Also, if it's
# a directory, it might be the model directory, and tokenizer.model might
# be in the parent of that.
if path.is_dir():
vocab_file = "tokenizer.model"
if vocabtype == 'bpe':
vocab_file = "vocab.json"
path2 = path / vocab_file
# Use `.parent` instead of /.. to handle the symlink case better.
path3 = path.parent / vocab_file
if path2.exists():
path = path2
elif path3.exists():
path = path3
else:
raise FileNotFoundError(
f"Could not find {vocab_file} in {path} or its parent; "
"if it's in another directory, pass the directory as --vocab-dir")
def find_vocab_file_path(path: Path, vocab_file: str) -> Optional[Path]:
path2 = path / vocab_file
# Use `.parent` instead of /.. to handle the symlink case better.
path3 = path.parent / vocab_file
print(f"Loading vocab file '{path}', type '{vocabtype}'")
if path2.exists():
return path2
if path3.exists():
return path3
added_tokens_path = path.parent / "added_tokens.json"
if vocabtype == "bpe":
return BpeVocab(path, added_tokens_path if added_tokens_path.exists() else None)
elif vocabtype == "spm":
return SentencePieceVocab(path, added_tokens_path if added_tokens_path.exists() else None)
else:
raise ValueError(f"Unsupported vocabulary type {vocabtype}")
return None
def default_outfile(model_paths: list[Path], file_type: GGMLFileType) -> Path:
@ -1139,19 +1187,33 @@ def main(args_in: list[str] | None = None) -> None:
# We currently only support Q8_0 output on little endian systems.
output_choices.append("q8_0")
parser = argparse.ArgumentParser(description="Convert a LLaMa model to a GGML compatible file")
parser.add_argument("--awq-path", type=Path, help="Path to scale awq cache file", default=None)
parser.add_argument("--dump", action="store_true", help="don't convert, just show what's in the model")
parser.add_argument("--dump-single", action="store_true", help="don't convert, just show what's in a single model file")
parser.add_argument("--vocab-only", action="store_true", help="extract only the vocab")
parser.add_argument("--outtype", choices=output_choices, help="output format - note: q8_0 may be very slow (default: f16 or f32 based on input)")
parser.add_argument("--vocab-dir", type=Path, help="directory containing tokenizer.model, if separate from model file")
parser.add_argument("--outfile", type=Path, help="path to write to; default: based on input")
parser.add_argument("model", type=Path, help="directory containing model file, or model file itself (*.pth, *.pt, *.bin, *.safetensors)")
parser.add_argument("--vocabtype", choices=["spm", "bpe"], help="vocab format (default: spm)", default="spm")
parser.add_argument("model", type=Path, help="directory containing model file, or model file itself (*.pth, *.pt, *.bin)")
parser.add_argument("--ctx", type=int, help="model training context (default: based on input)")
parser.add_argument("--concurrency", type=int, help=f"concurrency used for conversion (default: {DEFAULT_CONCURRENCY})", default = DEFAULT_CONCURRENCY)
parser.add_argument("--bigendian", action="store_true", help="model is executed on big endian machine")
parser.add_argument("--padvocab", action="store_true", help="add pad tokens when model vocab expects more than tokenizer metadata provides")
args = parser.parse_args(args_in)
if args.awq_path:
sys.path.insert(1, str(Path(__file__).parent / 'awq-py'))
from awq.apply_awq import add_scale_weights
tmp_model_path = args.model / "weighted_model"
if tmp_model_path.is_dir():
print(f"{tmp_model_path} exists as a weighted model.")
else:
tmp_model_path.mkdir(parents=True, exist_ok=True)
print("Saving new weighted model ...")
add_scale_weights(str(args.model), str(args.awq_path), str(tmp_model_path))
print(f"Saved weighted model at {tmp_model_path}.")
args.model = tmp_model_path
if args.dump_single:
model_plus = lazy_load_file(args.model)
do_dump_model(model_plus)
@ -1192,12 +1254,13 @@ def main(args_in: list[str] | None = None) -> None:
if not args.outfile:
raise ValueError("need --outfile if using --vocab-only")
# FIXME: Try to respect vocab_dir somehow?
vocab = load_vocab(args.vocab_dir or args.model, args.vocabtype)
vocab = VocabLoader(params, args.vocab_dir or args.model)
special_vocab = gguf.SpecialVocab(model_plus.paths[0].parent,
load_merges = args.vocabtype == 'bpe',
load_merges = True,
n_vocab = vocab.vocab_size)
outfile = args.outfile
OutputFile.write_vocab_only(outfile, params, vocab, special_vocab)
OutputFile.write_vocab_only(outfile, params, vocab, special_vocab,
endianess = endianess, pad_vocab = args.padvocab)
print(f"Wrote {outfile}")
return
@ -1205,12 +1268,15 @@ def main(args_in: list[str] | None = None) -> None:
vocab = model_plus.vocab
else:
vocab_dir = args.vocab_dir if args.vocab_dir else model_plus.paths[0].parent
vocab = load_vocab(vocab_dir, args.vocabtype)
vocab = VocabLoader(params, vocab_dir)
# FIXME: Try to respect vocab_dir somehow?
print(f"Vocab info: {vocab}")
special_vocab = gguf.SpecialVocab(model_plus.paths[0].parent,
load_merges = args.vocabtype == 'bpe',
load_merges = True,
n_vocab = vocab.vocab_size)
print(f"Special vocab info: {special_vocab}")
model = model_plus.model
model = convert_model_names(model, params)
ftype = pick_output_type(model, args.outtype)
@ -1220,7 +1286,8 @@ def main(args_in: list[str] | None = None) -> None:
params.ftype = ftype
print(f"Writing {outfile}, format {ftype}")
OutputFile.write_all(outfile, ftype, params, model, vocab, special_vocab, concurrency = args.concurrency, endianess=endianess)
OutputFile.write_all(outfile, ftype, params, model, vocab, special_vocab,
concurrency = args.concurrency, endianess = endianess, pad_vocab = args.padvocab)
print(f"Wrote {outfile}")

BIN
docs/llama-star/idea-arch.key Executable file
View file

Binary file not shown.

BIN
docs/llama-star/idea-arch.pdf Normal file
View file

Binary file not shown.

2
examples/CMakeLists.txt
View file

@ -32,6 +32,8 @@ else()
add_subdirectory(save-load-state)
add_subdirectory(simple)
add_subdirectory(speculative)
add_subdirectory(lookahead)
add_subdirectory(lookup)
add_subdirectory(train-text-from-scratch)
if (LLAMA_METAL)
add_subdirectory(metal)

33
examples/baby-llama/baby-llama.cpp
View file

@ -575,10 +575,7 @@ static struct ggml_tensor * forward(
// KQ_scaled = KQ / sqrt(n_embd/n_head)
// KQ_scaled shape [n_past + N, N, n_head, 1]
struct ggml_tensor * KQ_scaled =
ggml_scale(ctx0,
KQ,
ggml_new_f32(ctx0, 1.0f/sqrtf(float(n_embd)/n_head)));
struct ggml_tensor * KQ_scaled = ggml_scale(ctx0, KQ, 1.0f/sqrtf(float(n_embd)/n_head));
// KQ_masked = mask_past(KQ_scaled)
// KQ_masked shape [n_past + N, N, n_head, 1]
@ -844,10 +841,7 @@ static struct ggml_tensor * forward_batch(
// KQ_scaled = KQ / sqrt(n_embd/n_head)
// KQ_scaled shape [n_past + N, N, n_head, n_batch]
struct ggml_tensor * KQ_scaled =
ggml_scale(ctx0,
KQ,
ggml_new_f32(ctx0, 1.0f/sqrtf(float(n_embd)/n_head)));
struct ggml_tensor * KQ_scaled = ggml_scale(ctx0, KQ, 1.0f/sqrtf(float(n_embd)/n_head));
assert_shape_4d(KQ_scaled, n_past + N, N, n_head, n_batch);
// KQ_masked = mask_past(KQ_scaled)
@ -1131,10 +1125,7 @@ static struct ggml_tensor * forward_lora(
// KQ_scaled = KQ / sqrt(n_embd/n_head)
// KQ_scaled shape [n_past + N, N, n_head, 1]
struct ggml_tensor * KQ_scaled =
ggml_scale(ctx0,
KQ,
ggml_new_f32(ctx0, 1.0f/sqrtf(float(n_embd)/n_head)));
struct ggml_tensor * KQ_scaled = ggml_scale(ctx0, KQ, 1.0f/sqrtf(float(n_embd)/n_head));
// KQ_masked = mask_past(KQ_scaled)
// KQ_masked shape [n_past + N, N, n_head, 1]
@ -1258,9 +1249,9 @@ static struct ggml_tensor * forward_lora(
}
static void sample_softmax(struct ggml_tensor * logits, struct ggml_tensor * probs, struct ggml_tensor * best_samples) {
assert(logits->n_dims == 2);
assert(probs->n_dims == 2);
assert(best_samples->n_dims == 1);
assert(ggml_is_matrix(logits));
assert(ggml_is_matrix(probs));
assert(ggml_is_vector(best_samples));
assert(logits->ne[1] == best_samples->ne[0]);
assert(logits->ne[0] == probs->ne[0]);
assert(logits->ne[1] == probs->ne[1]);
@ -1292,9 +1283,9 @@ static void sample_softmax_batch(
struct ggml_context * ctx, struct ggml_tensor * logits, struct ggml_tensor * probs,
struct ggml_tensor * best_samples
) {
GGML_ASSERT(best_samples->n_dims == 2);
GGML_ASSERT(logits->n_dims == 3);
GGML_ASSERT(probs->n_dims == 3);
GGML_ASSERT(ggml_is_matrix(best_samples));
GGML_ASSERT(ggml_is_3d(logits));
GGML_ASSERT(ggml_is_3d(probs));
int n_tokens = best_samples->ne[0];
int n_batch = best_samples->ne[1];
int n_vocab = logits->ne[0];
@ -1334,7 +1325,7 @@ static void print_row(struct ggml_tensor * probs, int i) {
}
static void print_matrix(struct ggml_tensor * probs) {
assert(probs->n_dims == 2);
assert(ggml_is_matrix(probs));
for (int i = 0; i < probs->ne[1]; ++i) {
for (int k = 0; k < probs->ne[0]; ++k) {
float p = ggml_get_f32_1d(probs, i*probs->ne[0] + k);
@ -1386,8 +1377,8 @@ static void get_example_targets(int example_id, struct ggml_tensor * tokens_inpu
static void get_example_targets_batch(
struct ggml_context * ctx, int example_id, struct ggml_tensor * tokens_input, struct ggml_tensor * targets
) {
GGML_ASSERT(tokens_input->n_dims == 2);
GGML_ASSERT( targets->n_dims == 3);
GGML_ASSERT(ggml_is_matrix(tokens_input));
GGML_ASSERT(ggml_is_3d(targets));
int n_tokens = tokens_input->ne[0];
int n_batch = tokens_input->ne[1];
GGML_ASSERT(n_tokens == targets->ne[1]);

2
examples/batched-bench/batched-bench.cpp
View file

@ -155,7 +155,7 @@ int main(int argc, char ** argv) {
}
LOG_TEE("\n");
LOG_TEE("%s: n_kv_max = %d, is_pp_shared = %d, n_gpu_layers = %d, mmq = %d\n", __func__, n_kv_max, is_pp_shared, n_gpu_layers, mmq);
LOG_TEE("%s: n_kv_max = %d, is_pp_shared = %d, n_gpu_layers = %d, mmq = %d, n_threads = %d, n_threads_batch = %d\n", __func__, n_kv_max, is_pp_shared, n_gpu_layers, mmq, ctx_params.n_threads, ctx_params.n_threads_batch);
LOG_TEE("\n");
LOG_TEE("|%6s | %6s | %4s | %6s | %8s | %8s | %8s | %8s | %8s | %8s |\n", "PP", "TG", "B", "N_KV", "T_PP s", "S_PP t/s", "T_TG s", "S_TG t/s", "T s", "S t/s");

2
examples/batched.swift/README.md
View file

@ -1,4 +1,4 @@
This is a swift clone of `examples/batched`.
$ `make`
$ `./swift MODEL_PATH [PROMPT] [PARALLEL]`
$ `./batched_swift MODEL_PATH [PROMPT] [PARALLEL]`

17
examples/batched.swift/Sources/main.swift
View file

@ -153,7 +153,7 @@ while n_cur <= n_len {
// const llama_token new_token_id = llama_sample_token_greedy(ctx, &candidates_p);
// is it an end of stream? -> mark the stream as finished
if new_token_id == llama_token_eos(context) || n_cur == n_len {
if new_token_id == llama_token_eos(model) || n_cur == n_len {
i_batch[i] = -1
// print("")
if n_parallel > 1 {
@ -215,9 +215,10 @@ print("decoded \(n_decode) tokens in \(String(format: "%.2f", Double(t_main_end
llama_print_timings(context)
private func tokenize(text: String, add_bos: Bool) -> [llama_token] {
let n_tokens = text.count + (add_bos ? 1 : 0)
let utf8Count = text.utf8.count
let n_tokens = utf8Count + (add_bos ? 1 : 0)
let tokens = UnsafeMutablePointer<llama_token>.allocate(capacity: n_tokens)
let tokenCount = llama_tokenize(model, text, Int32(text.count), tokens, Int32(n_tokens), add_bos, /*special tokens*/ false)
let tokenCount = llama_tokenize(model, text, Int32(utf8Count), tokens, Int32(n_tokens), add_bos, /*special tokens*/ false)
var swiftTokens: [llama_token] = []
for i in 0 ..< tokenCount {
swiftTokens.append(tokens[Int(i)])
@ -230,18 +231,15 @@ private func token_to_piece(token: llama_token, buffer: inout [CChar]) -> String
var result = [CChar](repeating: 0, count: 8)
let nTokens = llama_token_to_piece(model, token, &result, Int32(result.count))
if nTokens < 0 {
if result.count >= -Int(nTokens) {
result.removeLast(-Int(nTokens))
} else {
result.removeAll()
}
let actualTokensCount = -Int(nTokens)
result = .init(repeating: 0, count: actualTokensCount)
let check = llama_token_to_piece(
model,
token,
&result,
Int32(result.count)
)
assert(check == nTokens)
assert(check == actualTokensCount)
} else {
result.removeLast(result.count - Int(nTokens))
}
@ -259,5 +257,4 @@ private func token_to_piece(token: llama_token, buffer: inout [CChar]) -> String
buffer = []
return bufferString
}
return nil
}

14
examples/benchmark/benchmark-matmult.cpp
View file

@ -129,13 +129,13 @@ int main(int argc, char ** argv) {
const ggml_type qtype = GGML_TYPE_Q4_1;
size_t ctx_size = 0;
ctx_size += sizex*sizey*ggml_type_sizef(GGML_TYPE_F32);
ctx_size += sizex*sizey*ggml_type_sizef(GGML_TYPE_F32);
ctx_size += sizex*sizez*ggml_type_sizef(GGML_TYPE_F32);
ctx_size += sizex*sizey*ggml_type_sizef(qtype);
ctx_size += sizex*sizey*ggml_type_sizef(qtype);
ctx_size += sizex*sizey*ggml_type_sizef(GGML_TYPE_F32); // BLAS
ctx_size += sizex*sizey*ggml_type_sizef(GGML_TYPE_F32); // BLAS
ctx_size += ggml_row_size(GGML_TYPE_F32, sizex*sizey);
ctx_size += ggml_row_size(GGML_TYPE_F32, sizex*sizey);
ctx_size += ggml_row_size(GGML_TYPE_F32, sizex*sizez);
ctx_size += ggml_row_size(qtype, sizex*sizey);
ctx_size += ggml_row_size(qtype, sizex*sizey);
ctx_size += ggml_row_size(GGML_TYPE_F32, sizex*sizey); // BLAS
ctx_size += ggml_row_size(GGML_TYPE_F32, sizex*sizey); // BLAS
ctx_size += 1024*1024*16;
printf("Allocating Memory of size %zi bytes, %zi MB\n",ctx_size, (ctx_size/1024/1024));

4
examples/convert-llama2c-to-ggml/convert-llama2c-to-ggml.cpp
View file

@ -427,7 +427,7 @@ static void print_row(struct ggml_tensor * probs, int i) {
}
static void print_matrix(struct ggml_tensor * probs) {
assert(probs->n_dims == 2);
assert(ggml_is_matrix(probs));
for (int i = 0; i < probs->ne[1]; ++i) {
for (int k = 0; k < probs->ne[0]; ++k) {
float p = get_f32_2d(probs, k, i);
@ -639,7 +639,7 @@ static void load_vocab(const char *filename, Config *config, struct llama_vocab
static void convert_weights_ak_to_gg(struct ggml_tensor * gg_weights, const float * karpathy_weights) {
int ct;
switch (gg_weights->n_dims){
switch (ggml_n_dims(gg_weights)) {
case 1:
ct = 0;
for (int i0 = 0; i0 < gg_weights->ne[0]; i0++){

2
examples/export-lora/export-lora.cpp
View file

@ -309,7 +309,7 @@ static struct ggml_cgraph * build_graph_lora(
) {
struct ggml_tensor * ab = ggml_mul_mat(ctx, lora_a, lora_b);
if (scaling != 1.0f) {
ab = ggml_scale(ctx, ab, ggml_new_f32(ctx, scaling));
ab = ggml_scale(ctx, ab, scaling);
}
struct ggml_tensor * res = ggml_add_inplace(ctx, tensor, ab);

73
examples/finetune/finetune.cpp
View file

@ -196,13 +196,13 @@ static const char * LLM_TENSOR_FFN_DOWN = "blk.%d.ffn_down";
static const char * LLM_TENSOR_FFN_UP = "blk.%d.ffn_up";
static void print_params(struct my_llama_hparams * params) {
printf("%s: n_vocab: %u\n", __func__, params->n_vocab);
printf("%s: n_ctx: %u\n", __func__, params->n_ctx);
printf("%s: n_embd: %u\n", __func__, params->n_embd);
printf("%s: n_ff: %u\n", __func__, params->n_ff);
printf("%s: n_head: %u\n", __func__, params->n_head);
printf("%s: n_head_kv: %u\n", __func__, params->n_head_kv);
printf("%s: n_layer: %u\n", __func__, params->n_layer);
printf("%s: n_vocab : %u\n", __func__, params->n_vocab);
printf("%s: n_ctx : %u\n", __func__, params->n_ctx);
printf("%s: n_embd : %u\n", __func__, params->n_embd);
printf("%s: n_ff : %u\n", __func__, params->n_ff);
printf("%s: n_head : %u\n", __func__, params->n_head);
printf("%s: n_head_kv : %u\n", __func__, params->n_head_kv);
printf("%s: n_layer : %u\n", __func__, params->n_layer);
printf("%s: norm_rms_eps : %f\n", __func__, params->f_norm_rms_eps);
printf("%s: rope_freq_base : %f\n", __func__, params->rope_freq_base);
printf("%s: rope_freq_scale : %f\n", __func__, params->rope_freq_scale);
@ -269,7 +269,7 @@ static void load_model_hparams_gguf(struct gguf_context * ctx, struct my_llama_h
float rope_freq_scale = 1.0f;
GGUF_GET_KEY(ctx, hparams->f_norm_rms_eps, gguf_get_val_f32, GGUF_TYPE_FLOAT32, false, kv(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS));
GGUF_GET_KEY(ctx, hparams->rope_freq_base, gguf_get_val_f32, GGUF_TYPE_FLOAT32, false, kv(LLM_KV_ROPE_FREQ_BASE));
GGUF_GET_KEY(ctx, rope_freq_scale, gguf_get_val_f32, GGUF_TYPE_FLOAT32, false, kv(LLM_KV_ROPE_SCALE_LINEAR));
GGUF_GET_KEY(ctx, rope_freq_scale, gguf_get_val_f32, GGUF_TYPE_FLOAT32, false, kv(LLM_KV_ROPE_SCALE_LINEAR));
if (rope_freq_scale != 1.0f) {
hparams->rope_freq_scale = 1.0f / rope_freq_scale;
}
@ -612,6 +612,7 @@ static struct ggml_tensor * llama_build_lora_finetune_graphs(
const int n_rot = hparams.n_embd_head();
const int n_embd_head = hparams.n_embd_head();
const int n_embd_gqa = hparams.n_embd_gqa();
const float rms_norm_eps = hparams.f_norm_rms_eps;
const float rope_freq_base = hparams.rope_freq_base;
const float rope_freq_scale = hparams.rope_freq_scale;
@ -680,10 +681,7 @@ static struct ggml_tensor * llama_build_lora_finetune_graphs(
checkpoints.push_back(t01);
}
struct ggml_tensor * kv_scale = NULL;
if (!enable_flash_attn) {
kv_scale = ggml_new_f32(ctx, 1.0f/sqrtf(float(n_embd)/n_head));
}
const float kv_scale = 1.0f/sqrtf(float(n_embd)/n_head);
for (int il = 0; il < n_layer; ++il) {
struct my_llama_layer & layer = model->layers[il];
@ -781,32 +779,32 @@ static struct ggml_tensor * llama_build_lora_finetune_graphs(
// make sure some tensors are not reallocated by inserting new temporary nodes depending on them
int n_leafs_before = gb->n_leafs;
int n_nodes_before = gb->n_nodes;
struct ggml_tensor * one = ggml_new_f32(ctx, 1.0f);
// output tensors
ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, t35, one));
ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, t36, one));
ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, t35, 1.0f));
ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, t36, 1.0f));
// input gradient
ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, t36->grad, one));
ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, t36->grad, 1.0f));
GGML_ASSERT(t36->grad->data == NULL && t36->grad->view_src == NULL);
ggml_allocr_alloc(alloc, t36->grad);
// KQ_pos
ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, KQ_pos, one));
ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, KQ_pos, 1.0f));
// make sure base model tensors data cannot be used in viewable operations
ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, model->tok_embeddings, one));
ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, model->norm, one));
ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, model->output, one));
ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, model->tok_embeddings, 1.0f));
ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, model->norm, 1.0f));
ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, model->output, 1.0f));
for (int il = 0; il < n_layer; ++il) {
struct my_llama_layer & layer = model->layers[il];
ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, layer.attention_norm, one));
ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, layer.ffn_norm, one));
ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, layer.wq, one));
ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, layer.wk, one));
ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, layer.wv, one));
ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, layer.wo, one));
ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, layer.w1, one));
ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, layer.w2, one));
ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, layer.w3, one));
ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, layer.attention_norm, 1.0f));
ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, layer.ffn_norm, 1.0f));
ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, layer.wq, 1.0f));
ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, layer.wk, 1.0f));
ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, layer.wv, 1.0f));
ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, layer.wo, 1.0f));
ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, layer.w1, 1.0f));
ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, layer.w2, 1.0f));
ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, layer.w3, 1.0f));
}
// allocating checkpoints in one block to reduce memory fragmentation
@ -1110,7 +1108,7 @@ static void write_tensor(struct llama_file * file, struct ggml_tensor * tensor,
name = ggml_get_name(tensor);
}
uint32_t name_len = strlen(name);
uint32_t nd = tensor->n_dims;
uint32_t nd = ggml_n_dims(tensor);
uint32_t ne[4] = { (uint32_t)tensor->ne[0],
(uint32_t)tensor->ne[1],
(uint32_t)tensor->ne[2],
@ -1620,8 +1618,6 @@ int main(int argc, char ** argv) {
opt->params.adam.gclip = params.common.adam_gclip;
opt->params.adam.eps_f = params.common.adam_eps_f;
ggml_allocr * alloc = NULL;
printf("%s: init model\n", __func__);
bool existed = load_checkpoint_lora_file(params.common.fn_checkpoint_in, &model, &lora, train);
@ -1725,10 +1721,9 @@ int main(int argc, char ** argv) {
// allocate input tensors
mem_input_data.resize(max_input_size);
alloc = ggml_allocr_new(mem_input_data.data(), mem_input_data.size(), tensor_alignment);
ggml_allocr_alloc(alloc, tokens_input);
ggml_allocr_alloc(alloc, target_probs);
ggml_allocr_free(alloc);
ggml_allocr_t alloc_inps = ggml_allocr_new(mem_input_data.data(), mem_input_data.size(), tensor_alignment);
ggml_allocr_alloc(alloc_inps, tokens_input);
ggml_allocr_alloc(alloc_inps, target_probs);
// context for compute tensors without their data
const size_t estimated_compute_size_wo_data = (
@ -1755,7 +1750,7 @@ int main(int argc, char ** argv) {
// find best evaluation order
for (unsigned order = 0; order < (unsigned) GGML_CGRAPH_EVAL_ORDER_COUNT; ++order) {
ctx_compute = ggml_init(ctx_compute_params);
alloc = ggml_allocr_new_measure(tensor_alignment);
ggml_allocr_t alloc = ggml_allocr_new_measure(tensor_alignment);
gf = ggml_new_graph_custom(ctx_compute, LLAMA_TRAIN_MAX_NODES, true);
gf->order = (enum ggml_cgraph_eval_order) order;
gb = ggml_new_graph_custom(ctx_compute, LLAMA_TRAIN_MAX_NODES, true);
@ -1788,7 +1783,7 @@ int main(int argc, char ** argv) {
// allocate compute tensors
mem_compute_data.resize(max_compute_size);
ctx_compute = ggml_init(ctx_compute_params);
alloc = ggml_allocr_new(mem_compute_data.data(), mem_compute_data.size(), tensor_alignment);
ggml_allocr_t alloc = ggml_allocr_new(mem_compute_data.data(), mem_compute_data.size(), tensor_alignment);
gf = ggml_new_graph_custom(ctx_compute, LLAMA_TRAIN_MAX_NODES, true);
gf->order = best_order;
gb = ggml_new_graph_custom(ctx_compute, LLAMA_TRAIN_MAX_NODES, true);
@ -1804,6 +1799,8 @@ int main(int argc, char ** argv) {
params.common.use_checkpointing
);
ggml_allocr_free(alloc);
ggml_allocr_free(alloc_inps);
// tokenize data
std::vector<llama_token> train_tokens;

2
examples/gguf/CMakeLists.txt
View file

@ -1,5 +1,5 @@
set(TARGET gguf)
add_executable(${TARGET} gguf.cpp)
install(TARGETS ${TARGET} RUNTIME)
target_link_libraries(${TARGET} PRIVATE llama ${CMAKE_THREAD_LIBS_INIT})
target_link_libraries(${TARGET} PRIVATE ggml ${CMAKE_THREAD_LIBS_INIT})
target_compile_features(${TARGET} PRIVATE cxx_std_11)

3
examples/gguf/gguf.cpp
View file

@ -1,5 +1,4 @@
#include "ggml.h"
#include "llama.h"
#include <cstdio>
#include <cinttypes>
@ -195,7 +194,7 @@ static bool gguf_ex_read_1(const std::string & fname) {
struct ggml_tensor * cur = ggml_get_tensor(ctx_data, name);
printf("%s: tensor[%d]: n_dims = %d, name = %s, data = %p\n", __func__, i, cur->n_dims, cur->name, cur->data);
printf("%s: tensor[%d]: n_dims = %d, name = %s, data = %p\n", __func__, i, ggml_n_dims(cur), cur->name, cur->data);
// print first 10 elements
const float * data = (const float *) cur->data;

113
examples/llama-bench/llama-bench.cpp
View file

@ -53,6 +53,13 @@ static std::vector<T> split(const std::string & str, char delim) {
return values;
}
template<typename T, typename F>
static std::vector<std::string> transform_to_str(const std::vector<T> & values, F f) {
std::vector<std::string> str_values;
std::transform(values.begin(), values.end(), std::back_inserter(str_values), f);
return str_values;
}
template<typename T>
static T avg(const std::vector<T> & v) {
if (v.empty()) {
@ -126,7 +133,8 @@ struct cmd_params {
std::vector<int> n_prompt;
std::vector<int> n_gen;
std::vector<int> n_batch;
std::vector<bool> f32_kv;
std::vector<ggml_type> type_k;
std::vector<ggml_type> type_v;
std::vector<int> n_threads;
std::vector<int> n_gpu_layers;
std::vector<int> main_gpu;
@ -142,7 +150,8 @@ static const cmd_params cmd_params_defaults = {
/* n_prompt */ {512},
/* n_gen */ {128},
/* n_batch */ {512},
/* f32_kv */ {false},
/* type_k */ {GGML_TYPE_F16},
/* type_v */ {GGML_TYPE_F16},
/* n_threads */ {get_num_physical_cores()},
/* n_gpu_layers */ {99},
/* main_gpu */ {0},
@ -162,7 +171,8 @@ static void print_usage(int /* argc */, char ** argv) {
printf(" -p, --n-prompt <n> (default: %s)\n", join(cmd_params_defaults.n_prompt, ",").c_str());
printf(" -n, --n-gen <n> (default: %s)\n", join(cmd_params_defaults.n_gen, ",").c_str());
printf(" -b, --batch-size <n> (default: %s)\n", join(cmd_params_defaults.n_batch, ",").c_str());
printf(" --memory-f32 <0|1> (default: %s)\n", join(cmd_params_defaults.f32_kv, ",").c_str());
printf(" -ctk <t>, --cache-type-k <t> (default: %s)\n", join(transform_to_str(cmd_params_defaults.type_k, ggml_type_name), ",").c_str());
printf(" -ctv <t>, --cache-type-v <t> (default: %s)\n", join(transform_to_str(cmd_params_defaults.type_v, ggml_type_name), ",").c_str());
printf(" -t, --threads <n> (default: %s)\n", join(cmd_params_defaults.n_threads, ",").c_str());
printf(" -ngl, --n-gpu-layers <n> (default: %s)\n", join(cmd_params_defaults.n_gpu_layers, ",").c_str());
printf(" -mg, --main-gpu <i> (default: %s)\n", join(cmd_params_defaults.main_gpu, ",").c_str());
@ -173,9 +183,32 @@ static void print_usage(int /* argc */, char ** argv) {
printf(" -v, --verbose (default: %s)\n", cmd_params_defaults.verbose ? "1" : "0");
printf("\n");
printf("Multiple values can be given for each parameter by separating them with ',' or by specifying the parameter multiple times.\n");
}
static ggml_type ggml_type_from_name(const std::string & s) {
if (s == "f16") {
return GGML_TYPE_F16;
}
if (s == "q8_0") {
return GGML_TYPE_Q8_0;
}
if (s == "q4_0") {
return GGML_TYPE_Q4_0;
}
if (s == "q4_1") {
return GGML_TYPE_Q4_1;
}
if (s == "q5_0") {
return GGML_TYPE_Q5_0;
}
if (s == "q5_1") {
return GGML_TYPE_Q5_1;
}
return GGML_TYPE_COUNT;
}
static cmd_params parse_cmd_params(int argc, char ** argv) {
cmd_params params;
std::string arg;
@ -224,13 +257,38 @@ static cmd_params parse_cmd_params(int argc, char ** argv) {
}
auto p = split<int>(argv[i], split_delim);
params.n_batch.insert(params.n_batch.end(), p.begin(), p.end());
} else if (arg == "--memory-f32") {
} else if (arg == "-ctk" || arg == "--cache-type-k") {
if (++i >= argc) {
invalid_param = true;
break;
}
auto p = split<int>(argv[i], split_delim);
params.f32_kv.insert(params.f32_kv.end(), p.begin(), p.end());
auto p = split<std::string>(argv[i], split_delim);
std::vector<ggml_type> types;
for (const auto & t : p) {
ggml_type gt = ggml_type_from_name(t);
if (gt == GGML_TYPE_COUNT) {
invalid_param = true;
break;
}
types.push_back(gt);
}
params.type_k.insert(params.type_k.end(), types.begin(), types.end());
} else if (arg == "-ctv" || arg == "--cache-type-v") {
if (++i >= argc) {
invalid_param = true;
break;
}
auto p = split<std::string>(argv[i], split_delim);
std::vector<ggml_type> types;
for (const auto & t : p) {
ggml_type gt = ggml_type_from_name(t);
if (gt == GGML_TYPE_COUNT) {
invalid_param = true;
break;
}
types.push_back(gt);
}
params.type_v.insert(params.type_v.end(), types.begin(), types.end());
} else if (arg == "-t" || arg == "--threads") {
if (++i >= argc) {
invalid_param = true;
@ -321,7 +379,8 @@ static cmd_params parse_cmd_params(int argc, char ** argv) {
if (params.n_prompt.empty()) { params.n_prompt = cmd_params_defaults.n_prompt; }
if (params.n_gen.empty()) { params.n_gen = cmd_params_defaults.n_gen; }
if (params.n_batch.empty()) { params.n_batch = cmd_params_defaults.n_batch; }
if (params.f32_kv.empty()) { params.f32_kv = cmd_params_defaults.f32_kv; }
if (params.type_k.empty()) { params.type_k = cmd_params_defaults.type_k; }
if (params.type_v.empty()) { params.type_v = cmd_params_defaults.type_v; }
if (params.n_gpu_layers.empty()) { params.n_gpu_layers = cmd_params_defaults.n_gpu_layers; }
if (params.main_gpu.empty()) { params.main_gpu = cmd_params_defaults.main_gpu; }
if (params.mul_mat_q.empty()) { params.mul_mat_q = cmd_params_defaults.mul_mat_q; }
@ -336,7 +395,8 @@ struct cmd_params_instance {
int n_prompt;
int n_gen;
int n_batch;
bool f32_kv;
ggml_type type_k;
ggml_type type_v;
int n_threads;
int n_gpu_layers;
int main_gpu;
@ -365,7 +425,8 @@ struct cmd_params_instance {
cparams.n_ctx = n_prompt + n_gen;
cparams.n_batch = n_batch;
cparams.f16_kv = !f32_kv;
cparams.type_k = type_k;
cparams.type_v = type_v;
cparams.mul_mat_q = mul_mat_q;
return cparams;
@ -380,7 +441,8 @@ static std::vector<cmd_params_instance> get_cmd_params_instances_int(const cmd_p
for (const auto & mg : params.main_gpu)
for (const auto & ts : params.tensor_split)
for (const auto & nb : params.n_batch)
for (const auto & fk : params.f32_kv)
for (const auto & tk : params.type_k)
for (const auto & tv : params.type_v)
for (const auto & mmq : params.mul_mat_q)
for (const auto & nt : params.n_threads) {
cmd_params_instance instance = {
@ -388,7 +450,8 @@ static std::vector<cmd_params_instance> get_cmd_params_instances_int(const cmd_p
/* .n_prompt = */ n_prompt,
/* .n_gen = */ n_gen,
/* .n_batch = */ nb,
/* .f32_kv = */ fk,
/* .type_k = */ tk,
/* .type_v = */ tv,
/* .n_threads = */ nt,
/* .n_gpu_layers = */ nl,
/* .main_gpu = */ mg,
@ -410,7 +473,8 @@ static std::vector<cmd_params_instance> get_cmd_params_instances(const cmd_param
for (const auto & mg : params.main_gpu)
for (const auto & ts : params.tensor_split)
for (const auto & nb : params.n_batch)
for (const auto & fk : params.f32_kv)
for (const auto & tk : params.type_k)
for (const auto & tv : params.type_v)
for (const auto & mmq : params.mul_mat_q)
for (const auto & nt : params.n_threads) {
for (const auto & n_prompt : params.n_prompt) {
@ -422,7 +486,8 @@ static std::vector<cmd_params_instance> get_cmd_params_instances(const cmd_param
/* .n_prompt = */ n_prompt,
/* .n_gen = */ 0,
/* .n_batch = */ nb,
/* .f32_kv = */ fk,
/* .type_k = */ tk,
/* .type_v = */ tv,
/* .n_threads = */ nt,
/* .n_gpu_layers = */ nl,
/* .main_gpu = */ mg,
@ -441,7 +506,8 @@ static std::vector<cmd_params_instance> get_cmd_params_instances(const cmd_param
/* .n_prompt = */ 0,
/* .n_gen = */ n_gen,
/* .n_batch = */ nb,
/* .f32_kv = */ fk,
/* .type_k = */ tk,
/* .type_v = */ tv,
/* .n_threads = */ nt,
/* .n_gpu_layers = */ nl,
/* .main_gpu = */ mg,
@ -489,7 +555,8 @@ struct test {
uint64_t model_n_params;
int n_batch;
int n_threads;
bool f32_kv;
ggml_type type_k;
ggml_type type_v;
int n_gpu_layers;
int main_gpu;
bool mul_mat_q;
@ -508,7 +575,8 @@ struct test {
model_n_params = llama_model_n_params(lmodel);
n_batch = inst.n_batch;
n_threads = inst.n_threads;
f32_kv = inst.f32_kv;
type_k = inst.type_k;
type_v = inst.type_v;
n_gpu_layers = inst.n_gpu_layers;
main_gpu = inst.main_gpu;
mul_mat_q = inst.mul_mat_q;
@ -571,7 +639,7 @@ struct test {
"cuda", "opencl", "metal", "gpu_blas", "blas",
"cpu_info", "gpu_info",
"model_filename", "model_type", "model_size", "model_n_params",
"n_batch", "n_threads", "f16_kv",
"n_batch", "n_threads", "type_k", "type_v",
"n_gpu_layers", "main_gpu", "mul_mat_q", "tensor_split",
"n_prompt", "n_gen", "test_time",
"avg_ns", "stddev_ns",
@ -621,7 +689,7 @@ struct test {
std::to_string(cuda), std::to_string(opencl), std::to_string(metal), std::to_string(gpu_blas), std::to_string(blas),
cpu_info, gpu_info,
model_filename, model_type, std::to_string(model_size), std::to_string(model_n_params),
std::to_string(n_batch), std::to_string(n_threads), std::to_string(!f32_kv),
std::to_string(n_batch), std::to_string(n_threads), ggml_type_name(type_k), ggml_type_name(type_v),
std::to_string(n_gpu_layers), std::to_string(main_gpu), std::to_string(mul_mat_q), tensor_split_str,
std::to_string(n_prompt), std::to_string(n_gen), test_time,
std::to_string(avg_ns()), std::to_string(stdev_ns()),
@ -805,8 +873,11 @@ struct markdown_printer : public printer {
if (params.n_batch.size() > 1 || params.n_batch != cmd_params_defaults.n_batch) {
fields.push_back("n_batch");
}
if (params.f32_kv.size() > 1 || params.f32_kv != cmd_params_defaults.f32_kv) {
fields.push_back("f16_kv");
if (params.type_k.size() > 1 || params.type_k != cmd_params_defaults.type_k) {
fields.push_back("type_k");
}
if (params.type_v.size() > 1 || params.type_v != cmd_params_defaults.type_v) {
fields.push_back("type_v");
}
if (params.main_gpu.size() > 1 || params.main_gpu != cmd_params_defaults.main_gpu) {
fields.push_back("main_gpu");

2
examples/llama.swiftui/.gitignore vendored Normal file
View file

@ -0,0 +1,2 @@
xcuserdata
xcshareddata

7
examples/llama.swiftui/README.md Normal file
View file

@ -0,0 +1,7 @@
# llama.swiftui
Local inference of llama.cpp on an iPhone.
So far I only tested with starcoder 1B model, but it can most likely handle 7B models as well.
https://github.com/bachittle/llama.cpp/assets/39804642/e290827a-4edb-4093-9642-2a5e399ec545

340
examples/llama.swiftui/llama.cpp.swift/LibLlama.swift Normal file
View file

@ -0,0 +1,340 @@
import Foundation
// import llama
enum LlamaError: Error {
case couldNotInitializeContext
}
func llama_batch_clear(_ batch: inout llama_batch) {
batch.n_tokens = 0
}
func llama_batch_add(_ batch: inout llama_batch, _ id: llama_token, _ pos: llama_pos, _ seq_ids: [llama_seq_id], _ logits: Bool) {
batch.token [Int(batch.n_tokens)] = id
batch.pos [Int(batch.n_tokens)] = pos
batch.n_seq_id[Int(batch.n_tokens)] = Int32(seq_ids.count)
for i in 0..<seq_ids.count {
batch.seq_id[Int(batch.n_tokens)]![Int(i)] = seq_ids[i]
}
batch.logits [Int(batch.n_tokens)] = logits ? 1 : 0
batch.n_tokens += 1
}
actor LlamaContext {
private var model: OpaquePointer
private var context: OpaquePointer
private var batch: llama_batch
private var tokens_list: [llama_token]
/// This variable is used to store temporarily invalid cchars
private var temporary_invalid_cchars: [CChar]
var n_len: Int32 = 64
var n_cur: Int32 = 0
var n_decode: Int32 = 0
init(model: OpaquePointer, context: OpaquePointer) {
self.model = model
self.context = context
self.tokens_list = []
self.batch = llama_batch_init(512, 0, 1)
self.temporary_invalid_cchars = []
}
deinit {
llama_batch_free(batch)
llama_free(context)
llama_free_model(model)
llama_backend_free()
}
static func create_context(path: String) throws -> LlamaContext {
llama_backend_init(false)
var model_params = llama_model_default_params()
#if targetEnvironment(simulator)
model_params.n_gpu_layers = 0
print("Running on simulator, force use n_gpu_layers = 0")
#endif
let model = llama_load_model_from_file(path, model_params)
guard let model else {
print("Could not load model at \(path)")
throw LlamaError.couldNotInitializeContext
}
let n_threads = max(1, min(8, ProcessInfo.processInfo.processorCount - 2))
print("Using \(n_threads) threads")
var ctx_params = llama_context_default_params()
ctx_params.seed = 1234
ctx_params.n_ctx = 2048
ctx_params.n_threads = UInt32(n_threads)
ctx_params.n_threads_batch = UInt32(n_threads)
let context = llama_new_context_with_model(model, ctx_params)
guard let context else {
print("Could not load context!")
throw LlamaError.couldNotInitializeContext
}
return LlamaContext(model: model, context: context)
}
func model_info() -> String {
let result = UnsafeMutablePointer<Int8>.allocate(capacity: 256)
result.initialize(repeating: Int8(0), count: 256)
defer {
result.deallocate()
}
// TODO: this is probably very stupid way to get the string from C
let nChars = llama_model_desc(model, result, 256)
let bufferPointer = UnsafeBufferPointer(start: result, count: Int(nChars))
var SwiftString = ""
for char in bufferPointer {
SwiftString.append(Character(UnicodeScalar(UInt8(char))))
}
return SwiftString
}
func get_n_tokens() -> Int32 {
return batch.n_tokens;
}
func completion_init(text: String) {
print("attempting to complete \"\(text)\"")
tokens_list = tokenize(text: text, add_bos: true)
temporary_invalid_cchars = []
let n_ctx = llama_n_ctx(context)
let n_kv_req = tokens_list.count + (Int(n_len) - tokens_list.count)
print("\n n_len = \(n_len), n_ctx = \(n_ctx), n_kv_req = \(n_kv_req)")
if n_kv_req > n_ctx {
print("error: n_kv_req > n_ctx, the required KV cache size is not big enough")
}
for id in tokens_list {
print(String(cString: token_to_piece(token: id) + [0]))
}
llama_batch_clear(&batch)
for i1 in 0..<tokens_list.count {
let i = Int(i1)
llama_batch_add(&batch, tokens_list[i], Int32(i), [0], false)
}
batch.logits[Int(batch.n_tokens) - 1] = 1 // true
if llama_decode(context, batch) != 0 {
print("llama_decode() failed")
}
n_cur = batch.n_tokens
}
func completion_loop() -> String {
var new_token_id: llama_token = 0
let n_vocab = llama_n_vocab(model)
let logits = llama_get_logits_ith(context, batch.n_tokens - 1)
var candidates = Array<llama_token_data>()
candidates.reserveCapacity(Int(n_vocab))
for token_id in 0..<n_vocab {
candidates.append(llama_token_data(id: token_id, logit: logits![Int(token_id)], p: 0.0))
}
candidates.withUnsafeMutableBufferPointer() { buffer in
var candidates_p = llama_token_data_array(data: buffer.baseAddress, size: buffer.count, sorted: false)
new_token_id = llama_sample_token_greedy(context, &candidates_p)
}
if new_token_id == llama_token_eos(context) || n_cur == n_len {
print("\n")
let new_token_str = String(cString: temporary_invalid_cchars + [0])
temporary_invalid_cchars.removeAll()
return new_token_str
}
let new_token_cchars = token_to_piece(token: new_token_id)
temporary_invalid_cchars.append(contentsOf: new_token_cchars)
let new_token_str: String
if let string = String(validatingUTF8: temporary_invalid_cchars + [0]) {
temporary_invalid_cchars.removeAll()
new_token_str = string
} else if (0 ..< temporary_invalid_cchars.count).contains(where: {$0 != 0 && String(validatingUTF8: Array(temporary_invalid_cchars.suffix($0)) + [0]) != nil}) {
// in this case, at least the suffix of the temporary_invalid_cchars can be interpreted as UTF8 string
let string = String(cString: temporary_invalid_cchars + [0])
temporary_invalid_cchars.removeAll()
new_token_str = string
} else {
new_token_str = ""
}
print(new_token_str)
// tokens_list.append(new_token_id)
llama_batch_clear(&batch)
llama_batch_add(&batch, new_token_id, n_cur, [0], true)
n_decode += 1
n_cur += 1
if llama_decode(context, batch) != 0 {
print("failed to evaluate llama!")
}
return new_token_str
}
func bench(pp: Int, tg: Int, pl: Int, nr: Int = 1) -> String {
var pp_avg: Double = 0
var tg_avg: Double = 0
var pp_std: Double = 0
var tg_std: Double = 0
for _ in 0..<nr {
// bench prompt processing
llama_batch_clear(&batch)
let n_tokens = pp
for i in 0..<n_tokens {
llama_batch_add(&batch, 0, Int32(i), [0], false)
}
batch.logits[Int(batch.n_tokens) - 1] = 1 // true
llama_kv_cache_clear(context)
let t_pp_start = ggml_time_us()
if llama_decode(context, batch) != 0 {
print("llama_decode() failed during prompt")
}
let t_pp_end = ggml_time_us()
// bench text generation
llama_kv_cache_clear(context)
let t_tg_start = ggml_time_us()
for i in 0..<tg {
llama_batch_clear(&batch)
for j in 0..<pl {
llama_batch_add(&batch, 0, Int32(i), [Int32(j)], true)
}
if llama_decode(context, batch) != 0 {
print("llama_decode() failed during text generation")
}
}
let t_tg_end = ggml_time_us()
llama_kv_cache_clear(context)
let t_pp = Double(t_pp_end - t_pp_start) / 1000000.0
let t_tg = Double(t_tg_end - t_tg_start) / 1000000.0
let speed_pp = Double(pp) / t_pp
let speed_tg = Double(pl*tg) / t_tg
pp_avg += speed_pp
tg_avg += speed_tg
pp_std += speed_pp * speed_pp
tg_std += speed_tg * speed_tg
print("pp \(speed_pp) t/s, tg \(speed_tg) t/s")
}
pp_avg /= Double(nr)
tg_avg /= Double(nr)
if nr > 1 {
pp_std = sqrt(pp_std / Double(nr - 1) - pp_avg * pp_avg * Double(nr) / Double(nr - 1))
tg_std = sqrt(tg_std / Double(nr - 1) - tg_avg * tg_avg * Double(nr) / Double(nr - 1))
} else {
pp_std = 0
tg_std = 0
}
let model_desc = model_info();
let model_size = String(format: "%.2f GiB", Double(llama_model_size(model)) / 1024.0 / 1024.0 / 1024.0);
let model_n_params = String(format: "%.2f B", Double(llama_model_n_params(model)) / 1e9);
let backend = "Metal";
let pp_avg_str = String(format: "%.2f", pp_avg);
let tg_avg_str = String(format: "%.2f", tg_avg);
let pp_std_str = String(format: "%.2f", pp_std);
let tg_std_str = String(format: "%.2f", tg_std);
var result = ""
result += String("| model | size | params | backend | test | t/s |\n")
result += String("| --- | --- | --- | --- | --- | --- |\n")
result += String("| \(model_desc) | \(model_size) | \(model_n_params) | \(backend) | pp \(pp) | \(pp_avg_str) ± \(pp_std_str) |\n")
result += String("| \(model_desc) | \(model_size) | \(model_n_params) | \(backend) | tg \(tg) | \(tg_avg_str) ± \(tg_std_str) |\n")
return result;
}
func clear() {
tokens_list.removeAll()
temporary_invalid_cchars.removeAll()
llama_kv_cache_clear(context)
}
private func tokenize(text: String, add_bos: Bool) -> [llama_token] {
let utf8Count = text.utf8.count
let n_tokens = utf8Count + (add_bos ? 1 : 0) + 1
let tokens = UnsafeMutablePointer<llama_token>.allocate(capacity: n_tokens)
let tokenCount = llama_tokenize(model, text, Int32(utf8Count), tokens, Int32(n_tokens), add_bos, false)
var swiftTokens: [llama_token] = []
for i in 0..<tokenCount {
swiftTokens.append(tokens[Int(i)])
}
tokens.deallocate()
return swiftTokens
}
/// - note: The result does not contain null-terminator
private func token_to_piece(token: llama_token) -> [CChar] {
let result = UnsafeMutablePointer<Int8>.allocate(capacity: 8)
result.initialize(repeating: Int8(0), count: 8)
defer {
result.deallocate()
}
let nTokens = llama_token_to_piece(model, token, result, 8)
if nTokens < 0 {
let newResult = UnsafeMutablePointer<Int8>.allocate(capacity: Int(-nTokens))
newResult.initialize(repeating: Int8(0), count: Int(-nTokens))
defer {
newResult.deallocate()
}
let nNewTokens = llama_token_to_piece(model, token, newResult, -nTokens)
let bufferPointer = UnsafeBufferPointer(start: newResult, count: Int(nNewTokens))
return Array(bufferPointer)
} else {
let bufferPointer = UnsafeBufferPointer(start: result, count: Int(nTokens))
return Array(bufferPointer)
}
}
}

5
examples/llama.swiftui/llama.cpp.swift/bridging-header.h Normal file
View file

@ -0,0 +1,5 @@
//
// Use this file to import your target's public headers that you would like to expose to Swift.
//
#import "llama.h"

483
examples/llama.swiftui/llama.swiftui.xcodeproj/project.pbxproj Normal file
View file

@ -0,0 +1,483 @@
// !$*UTF8*$!
{
archiveVersion = 1;
classes = {
};
objectVersion = 56;
objects = {
/* Begin PBXBuildFile section */
542376082B0D9BFB008E6A1C /* ggml-quants.c in Sources */ = {isa = PBXBuildFile; fileRef = 542376072B0D9BFB008E6A1C /* ggml-quants.c */; settings = {COMPILER_FLAGS = "-O3"; }; };
5423760B2B0D9C4B008E6A1C /* ggml-backend.c in Sources */ = {isa = PBXBuildFile; fileRef = 5423760A2B0D9C4B008E6A1C /* ggml-backend.c */; settings = {COMPILER_FLAGS = "-O3"; }; };
542378792ACE3F3500834A7B /* ggml-metal.metal in Resources */ = {isa = PBXBuildFile; fileRef = 549479C82AC9E10B00E0F78B /* ggml-metal.metal */; };
542EA09D2AC8723900A8AEE9 /* ggml.c in Sources */ = {isa = PBXBuildFile; fileRef = 542EA09B2AC8723900A8AEE9 /* ggml.c */; settings = {COMPILER_FLAGS = "-DGGML_USE_ACCELERATE -DGGML_USE_METAL -DGGML_USE_K_QUANTS -O3"; }; };
542EA0A02AC8725700A8AEE9 /* ggml-alloc.c in Sources */ = {isa = PBXBuildFile; fileRef = 542EA09F2AC8725700A8AEE9 /* ggml-alloc.c */; settings = {COMPILER_FLAGS = "-O3"; }; };
542EA0A32AC8729100A8AEE9 /* llama.cpp in Sources */ = {isa = PBXBuildFile; fileRef = 542EA0A12AC8729100A8AEE9 /* llama.cpp */; settings = {COMPILER_FLAGS = "-DGGML_USE_K_QUANTS -DGGML_USE_METAL -O3"; }; };
549479CB2AC9E16000E0F78B /* Metal.framework in Frameworks */ = {isa = PBXBuildFile; fileRef = 549479CA2AC9E16000E0F78B /* Metal.framework */; };
549479CD2AC9E42A00E0F78B /* ggml-metal.m in Sources */ = {isa = PBXBuildFile; fileRef = 549479C52AC9E0F200E0F78B /* ggml-metal.m */; settings = {COMPILER_FLAGS = "-fno-objc-arc -DGGML_SWIFT -DGGML_USE_METAL -O3"; }; };
7FA3D2B32B2EA2F600543F92 /* DownloadButton.swift in Sources */ = {isa = PBXBuildFile; fileRef = 7FA3D2B22B2EA2F600543F92 /* DownloadButton.swift */; };
8A1C83772AC328BD0096AF73 /* llama_swiftuiApp.swift in Sources */ = {isa = PBXBuildFile; fileRef = 8A1C83762AC328BD0096AF73 /* llama_swiftuiApp.swift */; };
8A1C83792AC328BD0096AF73 /* ContentView.swift in Sources */ = {isa = PBXBuildFile; fileRef = 8A1C83782AC328BD0096AF73 /* ContentView.swift */; };
8A1C837B2AC328BE0096AF73 /* Assets.xcassets in Resources */ = {isa = PBXBuildFile; fileRef = 8A1C837A2AC328BE0096AF73 /* Assets.xcassets */; };
8A1C837E2AC328BE0096AF73 /* Preview Assets.xcassets in Resources */ = {isa = PBXBuildFile; fileRef = 8A1C837D2AC328BE0096AF73 /* Preview Assets.xcassets */; };
8A39BE0A2AC7601100BFEB40 /* Accelerate.framework in Frameworks */ = {isa = PBXBuildFile; fileRef = 8A39BE092AC7601000BFEB40 /* Accelerate.framework */; };
8A3F84242AC4C891005E2EE8 /* models in Resources */ = {isa = PBXBuildFile; fileRef = 8A3F84232AC4C891005E2EE8 /* models */; };
8A907F332AC7138A006146EA /* LibLlama.swift in Sources */ = {isa = PBXBuildFile; fileRef = 8A907F322AC7134E006146EA /* LibLlama.swift */; };
8A9F7C4D2AC332EE008AE1EA /* LlamaState.swift in Sources */ = {isa = PBXBuildFile; fileRef = 8A9F7C4C2AC332EE008AE1EA /* LlamaState.swift */; };
/* End PBXBuildFile section */
/* Begin PBXFileReference section */
542376062B0D9BEA008E6A1C /* ggml-quants.h */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.h; name = "ggml-quants.h"; path = "../../ggml-quants.h"; sourceTree = "<group>"; };
542376072B0D9BFB008E6A1C /* ggml-quants.c */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.c; name = "ggml-quants.c"; path = "../../ggml-quants.c"; sourceTree = "<group>"; };
542376092B0D9C40008E6A1C /* ggml-backend.h */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.c.h; name = "ggml-backend.h"; path = "../../ggml-backend.h"; sourceTree = "<group>"; };
5423760A2B0D9C4B008E6A1C /* ggml-backend.c */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.c; name = "ggml-backend.c"; path = "../../ggml-backend.c"; sourceTree = "<group>"; };
542EA09B2AC8723900A8AEE9 /* ggml.c */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.c; name = ggml.c; path = ../../ggml.c; sourceTree = "<group>"; };
542EA09C2AC8723900A8AEE9 /* ggml.h */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.h; name = ggml.h; path = ../../ggml.h; sourceTree = "<group>"; };
542EA09E2AC8725700A8AEE9 /* ggml-alloc.h */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.h; name = "ggml-alloc.h"; path = "../../ggml-alloc.h"; sourceTree = "<group>"; };
542EA09F2AC8725700A8AEE9 /* ggml-alloc.c */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.c; name = "ggml-alloc.c"; path = "../../ggml-alloc.c"; sourceTree = "<group>"; };
542EA0A12AC8729100A8AEE9 /* llama.cpp */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.cpp.cpp; name = llama.cpp; path = ../../llama.cpp; sourceTree = "<group>"; };
542EA0A22AC8729100A8AEE9 /* llama.h */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.h; name = llama.h; path = ../../llama.h; sourceTree = "<group>"; };
549479C52AC9E0F200E0F78B /* ggml-metal.m */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.objc; name = "ggml-metal.m"; path = "../../ggml-metal.m"; sourceTree = "<group>"; };
549479C62AC9E0F200E0F78B /* ggml-metal.h */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.h; name = "ggml-metal.h"; path = "../../ggml-metal.h"; sourceTree = "<group>"; };
549479C82AC9E10B00E0F78B /* ggml-metal.metal */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.metal; name = "ggml-metal.metal"; path = "../../ggml-metal.metal"; sourceTree = "<group>"; };
549479CA2AC9E16000E0F78B /* Metal.framework */ = {isa = PBXFileReference; lastKnownFileType = wrapper.framework; name = Metal.framework; path = System/Library/Frameworks/Metal.framework; sourceTree = SDKROOT; };
7FA3D2B22B2EA2F600543F92 /* DownloadButton.swift */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.swift; path = DownloadButton.swift; sourceTree = "<group>"; };
8A08D20A2AC73B1500FE6CD4 /* bridging-header.h */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.c.h; path = "bridging-header.h"; sourceTree = "<group>"; };
8A1C83732AC328BD0096AF73 /* llama.swiftui.app */ = {isa = PBXFileReference; explicitFileType = wrapper.application; includeInIndex = 0; path = llama.swiftui.app; sourceTree = BUILT_PRODUCTS_DIR; };
8A1C83762AC328BD0096AF73 /* llama_swiftuiApp.swift */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.swift; path = llama_swiftuiApp.swift; sourceTree = "<group>"; };
8A1C83782AC328BD0096AF73 /* ContentView.swift */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.swift; path = ContentView.swift; sourceTree = "<group>"; };
8A1C837A2AC328BE0096AF73 /* Assets.xcassets */ = {isa = PBXFileReference; lastKnownFileType = folder.assetcatalog; path = Assets.xcassets; sourceTree = "<group>"; };
8A1C837D2AC328BE0096AF73 /* Preview Assets.xcassets */ = {isa = PBXFileReference; lastKnownFileType = folder.assetcatalog; path = "Preview Assets.xcassets"; sourceTree = "<group>"; };
8A39BE092AC7601000BFEB40 /* Accelerate.framework */ = {isa = PBXFileReference; lastKnownFileType = wrapper.framework; name = Accelerate.framework; path = System/Library/Frameworks/Accelerate.framework; sourceTree = SDKROOT; };
8A3F84232AC4C891005E2EE8 /* models */ = {isa = PBXFileReference; lastKnownFileType = folder; name = models; path = llama.swiftui/Resources/models; sourceTree = "<group>"; };
8A907F322AC7134E006146EA /* LibLlama.swift */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.swift; path = LibLlama.swift; sourceTree = "<group>"; };
8A9F7C4C2AC332EE008AE1EA /* LlamaState.swift */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.swift; path = LlamaState.swift; sourceTree = "<group>"; };
/* End PBXFileReference section */
/* Begin PBXFrameworksBuildPhase section */
8A1C83702AC328BD0096AF73 /* Frameworks */ = {
isa = PBXFrameworksBuildPhase;
buildActionMask = 2147483647;
files = (
549479CB2AC9E16000E0F78B /* Metal.framework in Frameworks */,
8A39BE0A2AC7601100BFEB40 /* Accelerate.framework in Frameworks */,
);
runOnlyForDeploymentPostprocessing = 0;
};
/* End PBXFrameworksBuildPhase section */
/* Begin PBXGroup section */
8A08D1F62AC7383900FE6CD4 /* llama.cpp */ = {
isa = PBXGroup;
children = (
5423760A2B0D9C4B008E6A1C /* ggml-backend.c */,
542376092B0D9C40008E6A1C /* ggml-backend.h */,
542376062B0D9BEA008E6A1C /* ggml-quants.h */,
542376072B0D9BFB008E6A1C /* ggml-quants.c */,
549479C82AC9E10B00E0F78B /* ggml-metal.metal */,
549479C62AC9E0F200E0F78B /* ggml-metal.h */,
549479C52AC9E0F200E0F78B /* ggml-metal.m */,
542EA09B2AC8723900A8AEE9 /* ggml.c */,
542EA09C2AC8723900A8AEE9 /* ggml.h */,
542EA09F2AC8725700A8AEE9 /* ggml-alloc.c */,
542EA09E2AC8725700A8AEE9 /* ggml-alloc.h */,
542EA0A12AC8729100A8AEE9 /* llama.cpp */,
542EA0A22AC8729100A8AEE9 /* llama.h */,
);
name = llama.cpp;
sourceTree = "<group>";
};
8A1C836A2AC328BD0096AF73 = {
isa = PBXGroup;
children = (
8A08D1F62AC7383900FE6CD4 /* llama.cpp */,
8A907F312AC7134E006146EA /* llama.cpp.swift */,
8A3F84232AC4C891005E2EE8 /* models */,
8A1C83752AC328BD0096AF73 /* llama.swiftui */,
8A1C83742AC328BD0096AF73 /* Products */,
8A39BE082AC7601000BFEB40 /* Frameworks */,
);
sourceTree = "<group>";
};
8A1C83742AC328BD0096AF73 /* Products */ = {
isa = PBXGroup;
children = (
8A1C83732AC328BD0096AF73 /* llama.swiftui.app */,
);
name = Products;
sourceTree = "<group>";
};
8A1C83752AC328BD0096AF73 /* llama.swiftui */ = {
isa = PBXGroup;
children = (
8A3F84102AC4BD85005E2EE8 /* Resources */,
8A9F7C4B2AC332DC008AE1EA /* Models */,
8A9F7C4A2AC332BF008AE1EA /* UI */,
8A1C83762AC328BD0096AF73 /* llama_swiftuiApp.swift */,
8A1C837A2AC328BE0096AF73 /* Assets.xcassets */,
8A1C837C2AC328BE0096AF73 /* Preview Content */,
);
path = llama.swiftui;
sourceTree = "<group>";
};
8A1C837C2AC328BE0096AF73 /* Preview Content */ = {
isa = PBXGroup;
children = (
8A1C837D2AC328BE0096AF73 /* Preview Assets.xcassets */,
);
path = "Preview Content";
sourceTree = "<group>";
};
8A39BE082AC7601000BFEB40 /* Frameworks */ = {
isa = PBXGroup;
children = (
549479CA2AC9E16000E0F78B /* Metal.framework */,
8A39BE092AC7601000BFEB40 /* Accelerate.framework */,
);
name = Frameworks;
sourceTree = "<group>";
};
8A3F84102AC4BD85005E2EE8 /* Resources */ = {
isa = PBXGroup;
children = (
8A3F84112AC4BD8C005E2EE8 /* models */,
);
path = Resources;
sourceTree = "<group>";
};
8A3F84112AC4BD8C005E2EE8 /* models */ = {
isa = PBXGroup;
children = (
);
path = models;
sourceTree = "<group>";
};
8A907F312AC7134E006146EA /* llama.cpp.swift */ = {
isa = PBXGroup;
children = (
8A08D20A2AC73B1500FE6CD4 /* bridging-header.h */,
8A907F322AC7134E006146EA /* LibLlama.swift */,
);
path = llama.cpp.swift;
sourceTree = "<group>";
};
8A9F7C4A2AC332BF008AE1EA /* UI */ = {
isa = PBXGroup;
children = (
7FA3D2B22B2EA2F600543F92 /* DownloadButton.swift */,
8A1C83782AC328BD0096AF73 /* ContentView.swift */,
);
path = UI;
sourceTree = "<group>";
};
8A9F7C4B2AC332DC008AE1EA /* Models */ = {
isa = PBXGroup;
children = (
8A9F7C4C2AC332EE008AE1EA /* LlamaState.swift */,
);
path = Models;
sourceTree = "<group>";
};
/* End PBXGroup section */
/* Begin PBXNativeTarget section */
8A1C83722AC328BD0096AF73 /* llama.swiftui */ = {
isa = PBXNativeTarget;
buildConfigurationList = 8A1C83812AC328BE0096AF73 /* Build configuration list for PBXNativeTarget "llama.swiftui" */;
buildPhases = (
8A1C836F2AC328BD0096AF73 /* Sources */,
8A1C83702AC328BD0096AF73 /* Frameworks */,
8A1C83712AC328BD0096AF73 /* Resources */,
);
buildRules = (
);
dependencies = (
);
name = llama.swiftui;
packageProductDependencies = (
);
productName = llama.swiftui;
productReference = 8A1C83732AC328BD0096AF73 /* llama.swiftui.app */;
productType = "com.apple.product-type.application";
};
/* End PBXNativeTarget section */
/* Begin PBXProject section */
8A1C836B2AC328BD0096AF73 /* Project object */ = {
isa = PBXProject;
attributes = {
BuildIndependentTargetsInParallel = 1;
LastSwiftUpdateCheck = 1500;
LastUpgradeCheck = 1500;
TargetAttributes = {
8A1C83722AC328BD0096AF73 = {
CreatedOnToolsVersion = 15.0;
LastSwiftMigration = 1500;
};
};
};
buildConfigurationList = 8A1C836E2AC328BD0096AF73 /* Build configuration list for PBXProject "llama.swiftui" */;
compatibilityVersion = "Xcode 14.0";
developmentRegion = en;
hasScannedForEncodings = 0;
knownRegions = (
en,
Base,
);
mainGroup = 8A1C836A2AC328BD0096AF73;
packageReferences = (
);
productRefGroup = 8A1C83742AC328BD0096AF73 /* Products */;
projectDirPath = "";
projectRoot = "";
targets = (
8A1C83722AC328BD0096AF73 /* llama.swiftui */,
);
};
/* End PBXProject section */
/* Begin PBXResourcesBuildPhase section */
8A1C83712AC328BD0096AF73 /* Resources */ = {
isa = PBXResourcesBuildPhase;
buildActionMask = 2147483647;
files = (
542378792ACE3F3500834A7B /* ggml-metal.metal in Resources */,
8A3F84242AC4C891005E2EE8 /* models in Resources */,
8A1C837E2AC328BE0096AF73 /* Preview Assets.xcassets in Resources */,
8A1C837B2AC328BE0096AF73 /* Assets.xcassets in Resources */,
);
runOnlyForDeploymentPostprocessing = 0;
};
/* End PBXResourcesBuildPhase section */
/* Begin PBXSourcesBuildPhase section */
8A1C836F2AC328BD0096AF73 /* Sources */ = {
isa = PBXSourcesBuildPhase;
buildActionMask = 2147483647;
files = (
542376082B0D9BFB008E6A1C /* ggml-quants.c in Sources */,
549479CD2AC9E42A00E0F78B /* ggml-metal.m in Sources */,
542EA09D2AC8723900A8AEE9 /* ggml.c in Sources */,
8A907F332AC7138A006146EA /* LibLlama.swift in Sources */,
542EA0A32AC8729100A8AEE9 /* llama.cpp in Sources */,
8A9F7C4D2AC332EE008AE1EA /* LlamaState.swift in Sources */,
8A1C83792AC328BD0096AF73 /* ContentView.swift in Sources */,
8A1C83772AC328BD0096AF73 /* llama_swiftuiApp.swift in Sources */,
7FA3D2B32B2EA2F600543F92 /* DownloadButton.swift in Sources */,
542EA0A02AC8725700A8AEE9 /* ggml-alloc.c in Sources */,
5423760B2B0D9C4B008E6A1C /* ggml-backend.c in Sources */,
);
runOnlyForDeploymentPostprocessing = 0;
};
/* End PBXSourcesBuildPhase section */
/* Begin XCBuildConfiguration section */
8A1C837F2AC328BE0096AF73 /* Debug */ = {
isa = XCBuildConfiguration;
buildSettings = {
ALWAYS_SEARCH_USER_PATHS = NO;
ASSETCATALOG_COMPILER_GENERATE_SWIFT_ASSET_SYMBOL_EXTENSIONS = YES;
CLANG_ANALYZER_NONNULL = YES;
CLANG_ANALYZER_NUMBER_OBJECT_CONVERSION = YES_AGGRESSIVE;
CLANG_CXX_LANGUAGE_STANDARD = "gnu++20";
CLANG_ENABLE_MODULES = YES;
CLANG_ENABLE_OBJC_ARC = YES;
CLANG_ENABLE_OBJC_WEAK = YES;
CLANG_WARN_BLOCK_CAPTURE_AUTORELEASING = YES;
CLANG_WARN_BOOL_CONVERSION = YES;
CLANG_WARN_COMMA = YES;
CLANG_WARN_CONSTANT_CONVERSION = YES;
CLANG_WARN_DEPRECATED_OBJC_IMPLEMENTATIONS = YES;
CLANG_WARN_DIRECT_OBJC_ISA_USAGE = YES_ERROR;
CLANG_WARN_DOCUMENTATION_COMMENTS = YES;
CLANG_WARN_EMPTY_BODY = YES;
CLANG_WARN_ENUM_CONVERSION = YES;
CLANG_WARN_INFINITE_RECURSION = YES;
CLANG_WARN_INT_CONVERSION = YES;
CLANG_WARN_NON_LITERAL_NULL_CONVERSION = YES;
CLANG_WARN_OBJC_IMPLICIT_RETAIN_SELF = YES;
CLANG_WARN_OBJC_LITERAL_CONVERSION = YES;
CLANG_WARN_OBJC_ROOT_CLASS = YES_ERROR;
CLANG_WARN_QUOTED_INCLUDE_IN_FRAMEWORK_HEADER = YES;
CLANG_WARN_RANGE_LOOP_ANALYSIS = YES;
CLANG_WARN_STRICT_PROTOTYPES = YES;
CLANG_WARN_SUSPICIOUS_MOVE = YES;
CLANG_WARN_UNGUARDED_AVAILABILITY = YES_AGGRESSIVE;
CLANG_WARN_UNREACHABLE_CODE = YES;
CLANG_WARN__DUPLICATE_METHOD_MATCH = YES;
COPY_PHASE_STRIP = NO;
DEBUG_INFORMATION_FORMAT = dwarf;
ENABLE_STRICT_OBJC_MSGSEND = YES;
ENABLE_TESTABILITY = YES;
ENABLE_USER_SCRIPT_SANDBOXING = YES;
GCC_C_LANGUAGE_STANDARD = gnu17;
GCC_DYNAMIC_NO_PIC = NO;
GCC_NO_COMMON_BLOCKS = YES;
GCC_OPTIMIZATION_LEVEL = 0;
GCC_PREPROCESSOR_DEFINITIONS = (
"DEBUG=1",
"$(inherited)",
);
GCC_WARN_64_TO_32_BIT_CONVERSION = YES;
GCC_WARN_ABOUT_RETURN_TYPE = YES_ERROR;
GCC_WARN_UNDECLARED_SELECTOR = YES;
GCC_WARN_UNINITIALIZED_AUTOS = YES_AGGRESSIVE;
GCC_WARN_UNUSED_FUNCTION = YES;
GCC_WARN_UNUSED_VARIABLE = YES;
IPHONEOS_DEPLOYMENT_TARGET = 17.0;
LOCALIZATION_PREFERS_STRING_CATALOGS = YES;
MTL_ENABLE_DEBUG_INFO = INCLUDE_SOURCE;
MTL_FAST_MATH = YES;
ONLY_ACTIVE_ARCH = YES;
SDKROOT = iphoneos;
SWIFT_ACTIVE_COMPILATION_CONDITIONS = "DEBUG $(inherited)";
SWIFT_OPTIMIZATION_LEVEL = "-Onone";
};
name = Debug;
};
8A1C83802AC328BE0096AF73 /* Release */ = {
isa = XCBuildConfiguration;
buildSettings = {
ALWAYS_SEARCH_USER_PATHS = NO;
ASSETCATALOG_COMPILER_GENERATE_SWIFT_ASSET_SYMBOL_EXTENSIONS = YES;
CLANG_ANALYZER_NONNULL = YES;
CLANG_ANALYZER_NUMBER_OBJECT_CONVERSION = YES_AGGRESSIVE;
CLANG_CXX_LANGUAGE_STANDARD = "gnu++20";
CLANG_ENABLE_MODULES = YES;
CLANG_ENABLE_OBJC_ARC = YES;
CLANG_ENABLE_OBJC_WEAK = YES;
CLANG_WARN_BLOCK_CAPTURE_AUTORELEASING = YES;
CLANG_WARN_BOOL_CONVERSION = YES;
CLANG_WARN_COMMA = YES;
CLANG_WARN_CONSTANT_CONVERSION = YES;
CLANG_WARN_DEPRECATED_OBJC_IMPLEMENTATIONS = YES;
CLANG_WARN_DIRECT_OBJC_ISA_USAGE = YES_ERROR;
CLANG_WARN_DOCUMENTATION_COMMENTS = YES;
CLANG_WARN_EMPTY_BODY = YES;
CLANG_WARN_ENUM_CONVERSION = YES;
CLANG_WARN_INFINITE_RECURSION = YES;
CLANG_WARN_INT_CONVERSION = YES;
CLANG_WARN_NON_LITERAL_NULL_CONVERSION = YES;
CLANG_WARN_OBJC_IMPLICIT_RETAIN_SELF = YES;
CLANG_WARN_OBJC_LITERAL_CONVERSION = YES;
CLANG_WARN_OBJC_ROOT_CLASS = YES_ERROR;
CLANG_WARN_QUOTED_INCLUDE_IN_FRAMEWORK_HEADER = YES;
CLANG_WARN_RANGE_LOOP_ANALYSIS = YES;
CLANG_WARN_STRICT_PROTOTYPES = YES;
CLANG_WARN_SUSPICIOUS_MOVE = YES;
CLANG_WARN_UNGUARDED_AVAILABILITY = YES_AGGRESSIVE;
CLANG_WARN_UNREACHABLE_CODE = YES;
CLANG_WARN__DUPLICATE_METHOD_MATCH = YES;
COPY_PHASE_STRIP = NO;
DEBUG_INFORMATION_FORMAT = "dwarf-with-dsym";
ENABLE_NS_ASSERTIONS = NO;
ENABLE_STRICT_OBJC_MSGSEND = YES;
ENABLE_USER_SCRIPT_SANDBOXING = YES;
GCC_C_LANGUAGE_STANDARD = gnu17;
GCC_NO_COMMON_BLOCKS = YES;
GCC_WARN_64_TO_32_BIT_CONVERSION = YES;
GCC_WARN_ABOUT_RETURN_TYPE = YES_ERROR;
GCC_WARN_UNDECLARED_SELECTOR = YES;
GCC_WARN_UNINITIALIZED_AUTOS = YES_AGGRESSIVE;
GCC_WARN_UNUSED_FUNCTION = YES;
GCC_WARN_UNUSED_VARIABLE = YES;
IPHONEOS_DEPLOYMENT_TARGET = 17.0;
LOCALIZATION_PREFERS_STRING_CATALOGS = YES;
MTL_ENABLE_DEBUG_INFO = NO;
MTL_FAST_MATH = YES;
SDKROOT = iphoneos;
SWIFT_COMPILATION_MODE = wholemodule;
VALIDATE_PRODUCT = YES;
};
name = Release;
};
8A1C83822AC328BE0096AF73 /* Debug */ = {
isa = XCBuildConfiguration;
buildSettings = {
ASSETCATALOG_COMPILER_APPICON_NAME = AppIcon;
ASSETCATALOG_COMPILER_GLOBAL_ACCENT_COLOR_NAME = AccentColor;
CLANG_ENABLE_MODULES = YES;
CODE_SIGN_STYLE = Automatic;
CURRENT_PROJECT_VERSION = 1;
DEVELOPMENT_ASSET_PATHS = "\"llama.swiftui/Preview Content\"";
DEVELOPMENT_TEAM = STLSG3FG8Q;
ENABLE_PREVIEWS = YES;
GENERATE_INFOPLIST_FILE = YES;
INFOPLIST_KEY_UIApplicationSceneManifest_Generation = YES;
INFOPLIST_KEY_UIApplicationSupportsIndirectInputEvents = YES;
INFOPLIST_KEY_UILaunchScreen_Generation = YES;
INFOPLIST_KEY_UISupportedInterfaceOrientations_iPad = "UIInterfaceOrientationPortrait UIInterfaceOrientationPortraitUpsideDown UIInterfaceOrientationLandscapeLeft UIInterfaceOrientationLandscapeRight";
INFOPLIST_KEY_UISupportedInterfaceOrientations_iPhone = "UIInterfaceOrientationPortrait UIInterfaceOrientationLandscapeLeft UIInterfaceOrientationLandscapeRight";
IPHONEOS_DEPLOYMENT_TARGET = 16.0;
LD_RUNPATH_SEARCH_PATHS = (
"$(inherited)",
"@executable_path/Frameworks",
);
MARKETING_VERSION = 1.0;
PRODUCT_BUNDLE_IDENTIFIER = "com.bachittle.llama-swift";
PRODUCT_NAME = "$(TARGET_NAME)";
SWIFT_EMIT_LOC_STRINGS = YES;
SWIFT_OBJC_BRIDGING_HEADER = "llama.cpp.swift/bridging-header.h";
SWIFT_OPTIMIZATION_LEVEL = "-Onone";
SWIFT_VERSION = 5.0;
TARGETED_DEVICE_FAMILY = "1,2";
};
name = Debug;
};
8A1C83832AC328BE0096AF73 /* Release */ = {
isa = XCBuildConfiguration;
buildSettings = {
ASSETCATALOG_COMPILER_APPICON_NAME = AppIcon;
ASSETCATALOG_COMPILER_GLOBAL_ACCENT_COLOR_NAME = AccentColor;
CLANG_ENABLE_MODULES = YES;
CODE_SIGN_STYLE = Automatic;
CURRENT_PROJECT_VERSION = 1;
DEVELOPMENT_ASSET_PATHS = "\"llama.swiftui/Preview Content\"";
DEVELOPMENT_TEAM = STLSG3FG8Q;
ENABLE_PREVIEWS = YES;
GENERATE_INFOPLIST_FILE = YES;
INFOPLIST_KEY_UIApplicationSceneManifest_Generation = YES;
INFOPLIST_KEY_UIApplicationSupportsIndirectInputEvents = YES;
INFOPLIST_KEY_UILaunchScreen_Generation = YES;
INFOPLIST_KEY_UISupportedInterfaceOrientations_iPad = "UIInterfaceOrientationPortrait UIInterfaceOrientationPortraitUpsideDown UIInterfaceOrientationLandscapeLeft UIInterfaceOrientationLandscapeRight";
INFOPLIST_KEY_UISupportedInterfaceOrientations_iPhone = "UIInterfaceOrientationPortrait UIInterfaceOrientationLandscapeLeft UIInterfaceOrientationLandscapeRight";
IPHONEOS_DEPLOYMENT_TARGET = 16.0;
LD_RUNPATH_SEARCH_PATHS = (
"$(inherited)",
"@executable_path/Frameworks",
);
MARKETING_VERSION = 1.0;
PRODUCT_BUNDLE_IDENTIFIER = "com.bachittle.llama-swift";
PRODUCT_NAME = "$(TARGET_NAME)";
SWIFT_EMIT_LOC_STRINGS = YES;
SWIFT_OBJC_BRIDGING_HEADER = "llama.cpp.swift/bridging-header.h";
SWIFT_VERSION = 5.0;
TARGETED_DEVICE_FAMILY = "1,2";
};
name = Release;
};
/* End XCBuildConfiguration section */
/* Begin XCConfigurationList section */
8A1C836E2AC328BD0096AF73 /* Build configuration list for PBXProject "llama.swiftui" */ = {
isa = XCConfigurationList;
buildConfigurations = (
8A1C837F2AC328BE0096AF73 /* Debug */,
8A1C83802AC328BE0096AF73 /* Release */,
);
defaultConfigurationIsVisible = 0;
defaultConfigurationName = Release;
};
8A1C83812AC328BE0096AF73 /* Build configuration list for PBXNativeTarget "llama.swiftui" */ = {
isa = XCConfigurationList;
buildConfigurations = (
8A1C83822AC328BE0096AF73 /* Debug */,
8A1C83832AC328BE0096AF73 /* Release */,
);
defaultConfigurationIsVisible = 0;
defaultConfigurationName = Release;
};
/* End XCConfigurationList section */
};
rootObject = 8A1C836B2AC328BD0096AF73 /* Project object */;
}

7
examples/llama.swiftui/llama.swiftui.xcodeproj/project.xcworkspace/contents.xcworkspacedata generated Normal file
View file

@ -0,0 +1,7 @@
<?xml version="1.0" encoding="UTF-8"?>
<Workspace
version = "1.0">
<FileRef
location = "self:">
</FileRef>
</Workspace>

8
examples/llama.swiftui/llama.swiftui.xcodeproj/project.xcworkspace/xcshareddata/IDEWorkspaceChecks.plist Normal file
View file

@ -0,0 +1,8 @@
<?xml version="1.0" encoding="UTF-8"?>
<!DOCTYPE plist PUBLIC "-//Apple//DTD PLIST 1.0//EN" "http://www.apple.com/DTDs/PropertyList-1.0.dtd">
<plist version="1.0">
<dict>
<key>IDEDidComputeMac32BitWarning</key>
<true/>
</dict>
</plist>

11
examples/llama.swiftui/llama.swiftui/Assets.xcassets/AccentColor.colorset/Contents.json Normal file
View file

@ -0,0 +1,11 @@
{
"colors" : [
{
"idiom" : "universal"
}
],
"info" : {
"author" : "xcode",
"version" : 1
}
}

13
examples/llama.swiftui/llama.swiftui/Assets.xcassets/AppIcon.appiconset/Contents.json Normal file
View file

@ -0,0 +1,13 @@
{
"images" : [
{
"idiom" : "universal",
"platform" : "ios",
"size" : "1024x1024"
}
],
"info" : {
"author" : "xcode",
"version" : 1
}
}

6
examples/llama.swiftui/llama.swiftui/Assets.xcassets/Contents.json Normal file
View file

@ -0,0 +1,6 @@
{
"info" : {
"author" : "xcode",
"version" : 1
}
}

85
examples/llama.swiftui/llama.swiftui/Models/LlamaState.swift Normal file
View file

@ -0,0 +1,85 @@
import Foundation
@MainActor
class LlamaState: ObservableObject {
@Published var messageLog = ""
@Published var cacheCleared = false
private var llamaContext: LlamaContext?
private var defaultModelUrl: URL? {
Bundle.main.url(forResource: "ggml-model", withExtension: "gguf", subdirectory: "models")
// Bundle.main.url(forResource: "llama-2-7b-chat", withExtension: "Q2_K.gguf", subdirectory: "models")
}
init() {
do {
try loadModel(modelUrl: defaultModelUrl)
} catch {
messageLog += "Error!\n"
}
}
func loadModel(modelUrl: URL?) throws {
messageLog += "Loading model...\n"
if let modelUrl {
llamaContext = try LlamaContext.create_context(path: modelUrl.path())
messageLog += "Loaded model \(modelUrl.lastPathComponent)\n"
} else {
messageLog += "Could not locate model\n"
}
}
func complete(text: String) async {
guard let llamaContext else {
return
}
await llamaContext.completion_init(text: text)
messageLog += "\(text)"
while await llamaContext.n_cur <= llamaContext.n_len {
let result = await llamaContext.completion_loop()
messageLog += "\(result)"
}
await llamaContext.clear()
messageLog += "\n\ndone\n"
}
func bench() async {
guard let llamaContext else {
return
}
messageLog += "\n"
messageLog += "Running benchmark...\n"
messageLog += "Model info: "
messageLog += await llamaContext.model_info() + "\n"
let t_start = DispatchTime.now().uptimeNanoseconds
await llamaContext.bench(pp: 8, tg: 4, pl: 1) // heat up
let t_end = DispatchTime.now().uptimeNanoseconds
let t_heat = Double(t_end - t_start) / 1_000_000_000.0
messageLog += "Heat up time: \(t_heat) seconds, please wait...\n"
// if more than 5 seconds, then we're probably running on a slow device
if t_heat > 5.0 {
messageLog += "Heat up time is too long, aborting benchmark\n"
return
}
let result = await llamaContext.bench(pp: 512, tg: 128, pl: 1, nr: 3)
messageLog += "\(result)"
messageLog += "\n"
}
func clear() async {
guard let llamaContext else {
return
}
await llamaContext.clear()
messageLog = ""
}
}

6
examples/llama.swiftui/llama.swiftui/Preview Content/Preview Assets.xcassets/Contents.json Normal file
View file

@ -0,0 +1,6 @@
{
"info" : {
"author" : "xcode",
"version" : 1
}
}

0
examples/llama.swiftui/llama.swiftui/Resources/models/.gitignore vendored Normal file
View file

161
examples/llama.swiftui/llama.swiftui/UI/ContentView.swift Normal file
View file

@ -0,0 +1,161 @@
import SwiftUI
struct ContentView: View {
@StateObject var llamaState = LlamaState()
@State private var multiLineText = ""
private static func cleanupModelCaches() {
// Delete all models (*.gguf)
let fileManager = FileManager.default
let documentsUrl = FileManager.default.urls(for: .documentDirectory, in: .userDomainMask)[0]
do {
let fileURLs = try fileManager.contentsOfDirectory(at: documentsUrl, includingPropertiesForKeys: nil)
for fileURL in fileURLs {
if fileURL.pathExtension == "gguf" {
try fileManager.removeItem(at: fileURL)
}
}
} catch {
print("Error while enumerating files \(documentsUrl.path): \(error.localizedDescription)")
}
}
var body: some View {
VStack {
ScrollView(.vertical, showsIndicators: true) {
Text(llamaState.messageLog)
.font(.system(size: 12))
.frame(maxWidth: .infinity, alignment: .leading)
.padding()
.onTapGesture {
UIApplication.shared.sendAction(#selector(UIResponder.resignFirstResponder), to: nil, from: nil, for: nil)
}
}
TextEditor(text: $multiLineText)
.frame(height: 80)
.padding()
.border(Color.gray, width: 0.5)
HStack {
Button("Send") {
sendText()
}
.padding(8)
.background(Color.blue)
.foregroundColor(.white)
.cornerRadius(8)
Button("Bench") {
bench()
}
.padding(8)
.background(Color.blue)
.foregroundColor(.white)
.cornerRadius(8)
Button("Clear") {
clear()
}
.padding(8)
.background(Color.blue)
.foregroundColor(.white)
.cornerRadius(8)
Button("Copy") {
UIPasteboard.general.string = llamaState.messageLog
}
.padding(8)
.background(Color.blue)
.foregroundColor(.white)
.cornerRadius(8)
}
VStack {
DownloadButton(
llamaState: llamaState,
modelName: "TinyLlama-1.1B (Q4_0, 0.6 GiB)",
modelUrl: "https://huggingface.co/TheBloke/TinyLlama-1.1B-1T-OpenOrca-GGUF/resolve/main/tinyllama-1.1b-1t-openorca.Q4_0.gguf?download=true",
filename: "tinyllama-1.1b-1t-openorca.Q4_0.gguf"
)
.font(.system(size: 12))
.padding(.top, 4)
.frame(maxWidth: .infinity, alignment: .leading)
DownloadButton(
llamaState: llamaState,
modelName: "TinyLlama-1.1B (Q8_0, 1.1 GiB)",
modelUrl: "https://huggingface.co/TheBloke/TinyLlama-1.1B-1T-OpenOrca-GGUF/resolve/main/tinyllama-1.1b-1t-openorca.Q8_0.gguf?download=true",
filename: "tinyllama-1.1b-1t-openorca.Q8_0.gguf"
)
.font(.system(size: 12))
DownloadButton(
llamaState: llamaState,
modelName: "TinyLlama-1.1B (F16, 2.2 GiB)",
modelUrl: "https://huggingface.co/ggml-org/models/resolve/main/tinyllama-1.1b/ggml-model-f16.gguf?download=true",
filename: "tinyllama-1.1b-f16.gguf"
)
.font(.system(size: 12))
.frame(maxWidth: .infinity, alignment: .leading)
DownloadButton(
llamaState: llamaState,
modelName: "Phi-2.7B (Q4_0, 1.6 GiB)",
modelUrl: "https://huggingface.co/ggml-org/models/resolve/main/phi-2/ggml-model-q4_0.gguf?download=true",
filename: "phi-2-q4_0.gguf"
)
.font(.system(size: 12))
DownloadButton(
llamaState: llamaState,
modelName: "Phi-2.7B (Q8_0, 2.8 GiB)",
modelUrl: "https://huggingface.co/ggml-org/models/resolve/main/phi-2/ggml-model-q8_0.gguf?download=true",
filename: "phi-2-q8_0.gguf"
)
.font(.system(size: 12))
.frame(maxWidth: .infinity, alignment: .leading)
DownloadButton(
llamaState: llamaState,
modelName: "Mistral-7B-v0.1 (Q4_0, 3.8 GiB)",
modelUrl: "https://huggingface.co/TheBloke/Mistral-7B-v0.1-GGUF/resolve/main/mistral-7b-v0.1.Q4_0.gguf?download=true",
filename: "mistral-7b-v0.1.Q4_0.gguf"
)
.font(.system(size: 12))
Button("Clear downloaded models") {
ContentView.cleanupModelCaches()
llamaState.cacheCleared = true
}
.padding(8)
.font(.system(size: 12))
}
}
.padding()
}
func sendText() {
Task {
await llamaState.complete(text: multiLineText)
multiLineText = ""
}
}
func bench() {
Task {
await llamaState.bench()
}
}
func clear() {
Task {
await llamaState.clear()
}
}
}
//#Preview {
// ContentView()
//}

122
examples/llama.swiftui/llama.swiftui/UI/DownloadButton.swift Normal file
View file

@ -0,0 +1,122 @@
import SwiftUI
struct DownloadButton: View {
@ObservedObject private var llamaState: LlamaState
private var modelName: String
private var modelUrl: String
private var filename: String
@State private var status: String
@State private var downloadTask: URLSessionDownloadTask?
@State private var progress = 0.0
@State private var observation: NSKeyValueObservation?
private static func getFileURL(filename: String) -> URL {
FileManager.default.urls(for: .documentDirectory, in: .userDomainMask)[0].appendingPathComponent(filename)
}
private func checkFileExistenceAndUpdateStatus() {
}
init(llamaState: LlamaState, modelName: String, modelUrl: String, filename: String) {
self.llamaState = llamaState
self.modelName = modelName
self.modelUrl = modelUrl
self.filename = filename
let fileURL = DownloadButton.getFileURL(filename: filename)
status = FileManager.default.fileExists(atPath: fileURL.path) ? "downloaded" : "download"
}
private func download() {
status = "downloading"
print("Downloading model \(modelName) from \(modelUrl)")
guard let url = URL(string: modelUrl) else { return }
let fileURL = DownloadButton.getFileURL(filename: filename)
downloadTask = URLSession.shared.downloadTask(with: url) { temporaryURL, response, error in
if let error = error {
print("Error: \(error.localizedDescription)")
return
}
guard let response = response as? HTTPURLResponse, (200...299).contains(response.statusCode) else {
print("Server error!")
return
}
do {
if let temporaryURL = temporaryURL {
try FileManager.default.copyItem(at: temporaryURL, to: fileURL)
print("Writing to \(filename) completed")
llamaState.cacheCleared = false
status = "downloaded"
}
} catch let err {
print("Error: \(err.localizedDescription)")
}
}
observation = downloadTask?.progress.observe(\.fractionCompleted) { progress, _ in
self.progress = progress.fractionCompleted
}
downloadTask?.resume()
}
var body: some View {
VStack {
if status == "download" {
Button(action: download) {
Text("Download " + modelName)
}
} else if status == "downloading" {
Button(action: {
downloadTask?.cancel()
status = "download"
}) {
Text("\(modelName) (Downloading \(Int(progress * 100))%)")
}
} else if status == "downloaded" {
Button(action: {
let fileURL = DownloadButton.getFileURL(filename: filename)
if !FileManager.default.fileExists(atPath: fileURL.path) {
download()
return
}
do {
try llamaState.loadModel(modelUrl: fileURL)
} catch let err {
print("Error: \(err.localizedDescription)")
}
}) {
Text("\(modelName) (Downloaded)")
}
} else {
Text("Unknown status")
}
}
.onDisappear() {
downloadTask?.cancel()
}
.onChange(of: llamaState.cacheCleared) { newValue in
if newValue {
downloadTask?.cancel()
let fileURL = DownloadButton.getFileURL(filename: filename)
status = FileManager.default.fileExists(atPath: fileURL.path) ? "downloaded" : "download"
}
}
}
}
// #Preview {
// DownloadButton(
// llamaState: LlamaState(),
// modelName: "TheBloke / TinyLlama-1.1B-1T-OpenOrca-GGUF (Q4_0)",
// modelUrl: "https://huggingface.co/TheBloke/TinyLlama-1.1B-1T-OpenOrca-GGUF/resolve/main/tinyllama-1.1b-1t-openorca.Q4_0.gguf?download=true",
// filename: "tinyllama-1.1b-1t-openorca.Q4_0.gguf"
// )
// }

10
examples/llama.swiftui/llama.swiftui/llama_swiftuiApp.swift Normal file
View file

@ -0,0 +1,10 @@
import SwiftUI
@main
struct llama_swiftuiApp: App {
var body: some Scene {
WindowGroup {
ContentView()
}
}
}

16
examples/llava/clip.cpp
View file

@ -330,12 +330,6 @@ static ggml_cgraph * clip_image_build_graph(const clip_ctx * ctx, const clip_ima
ggml_repeat(ctx0, model.pre_ln_b, embeddings));
}
struct ggml_tensor * KQ_scale = ggml_new_tensor_1d(ctx0, GGML_TYPE_F32, 1);
ggml_allocr_alloc(ctx->alloc, KQ_scale);
if (!ggml_allocr_is_measure(ctx->alloc)) {
ggml_set_f32(KQ_scale, 1.0f / sqrt((float)d_head));
}
// loop over layers
for (int il = 0; il < n_layer - 1; il++) {
struct ggml_tensor * cur = embeddings; // embeddings = residual, cur = hidden_states
@ -356,7 +350,7 @@ static ggml_cgraph * clip_image_build_graph(const clip_ctx * ctx, const clip_ima
struct ggml_tensor * Q =
ggml_add(ctx0, ggml_repeat(ctx0, model.layers[il].q_b, cur), ggml_mul_mat(ctx0, model.layers[il].q_w, cur));
Q = ggml_scale_inplace(ctx0, Q, KQ_scale);
Q = ggml_scale_inplace(ctx0, Q, 1.0f / sqrt((float)d_head));
Q = ggml_reshape_4d(ctx0, Q, d_head, n_head, num_positions, batch_size);
Q = ggml_cont(ctx0, ggml_permute(ctx0, Q, 0, 2, 1, 3));
Q = ggml_reshape_3d(ctx0, Q, d_head, num_positions, n_head * batch_size);
@ -514,7 +508,7 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
ctx_size += padded_size;
if (verbosity >= 3) {
printf("%s: tensor[%d]: n_dims = %d, name = %s, tensor_size=%zu, padded_size=%zu, offset=%zu\n", __func__, i,
cur->n_dims, cur->name, tensor_size, padded_size, offset);
ggml_n_dims(cur), cur->name, tensor_size, padded_size, offset);
}
}
}
@ -739,7 +733,7 @@ bool clip_image_preprocess(const clip_ctx * ctx, const clip_image_u8 * img, clip
temp->ny = longer_side;
temp->size = 3 * longer_side * longer_side;
temp->data = new uint8_t[temp->size]();
uint8_t bc[3] = {122, 116, 104}; // bakground color in RGB from LLaVA
uint8_t bc[3] = {122, 116, 104}; // background color in RGB from LLaVA
// fill with background color
for (size_t i = 0; i < temp->size; i++) {
@ -962,7 +956,7 @@ bool clip_model_quantize(const char * fname_inp, const char * fname_out, const i
}
// quantize only 2D tensors
quantize &= (cur->n_dims == 2);
quantize &= (ggml_n_dims(cur) == 2);
if (quantize) {
new_type = type;
@ -1035,7 +1029,7 @@ bool clip_model_quantize(const char * fname_inp, const char * fname_out, const i
fout.put(0);
}
printf("%s: n_dims = %d | quantize=%d | size = %f MB -> %f MB\n", name.c_str(), cur->n_dims, quantize,
printf("%s: n_dims = %d | quantize=%d | size = %f MB -> %f MB\n", name.c_str(), ggml_n_dims(cur), quantize,
orig_size / 1024.0 / 1024.0, new_size / 1024.0 / 1024.0);
}

54
examples/llava/convert-image-encoder-to-gguf.py
View file

@ -5,7 +5,7 @@ import json
import torch
import numpy as np
from gguf import *
from transformers import CLIPModel, CLIPProcessor
from transformers import CLIPModel, CLIPProcessor, CLIPVisionModel
TEXT = "clip.text"
VISION = "clip.vision"
@ -51,7 +51,7 @@ def bytes_to_unicode():
The reversible bpe codes work on unicode strings.
This means you need a large # of unicode characters in your vocab if you want to avoid UNKs.
When you're at something like a 10B token dataset you end up needing around 5K for decent coverage.
This is a signficant percentage of your normal, say, 32K bpe vocab.
This is a significant percentage of your normal, say, 32K bpe vocab.
To avoid that, we want lookup tables between utf-8 bytes and unicode strings.
And avoids mapping to whitespace/control characters the bpe code barfs on.
"""
@ -78,11 +78,19 @@ ap.add_argument("--text-only", action="store_true", required=False,
help="Save a text-only model. It can't be used to encode images")
ap.add_argument("--vision-only", action="store_true", required=False,
help="Save a vision-only model. It can't be used to encode texts")
ap.add_argument("--clip_model_is_vision", action="store_true", required=False,
help="The clip model is a pure vision model (ShareGPT4V vision extract for example)")
ap.add_argument("--llava-projector", help="Path to llava.projector file. If specified, save an image encoder for LLaVA models.")
ap.add_argument("--image-mean", nargs=3, type=float, required=False, help="Override image mean values")
ap.add_argument("--image-std", nargs=3, type=float, required=False, help="Override image std values")
ap.add_argument("-o", "--output-dir", help="Directory to save GGUF files. Default is the original model directory", default=None)
# Example --image_mean 0.48145466 0.4578275 0.40821073 --image_std 0.26862954 0.26130258 0.27577711
default_image_mean = [0.48145466, 0.4578275, 0.40821073]
default_image_std = [0.26862954, 0.26130258, 0.27577711]
ap.add_argument('--image_mean', type=float, nargs='+', help='Mean of the images for normalization (overrides processor) ', default=None)
ap.add_argument('--image_std', type=float, nargs='+', help='Standard deviation of the images for normalization (overrides processor)', default=None)
# with proper
args = ap.parse_args()
@ -96,15 +104,22 @@ if args.use_f32:
# output in the same directory as the model if output_dir is None
dir_model = args.model_dir
with open(dir_model + "/vocab.json", "r", encoding="utf-8") as f:
vocab = json.load(f)
tokens = [key for key in vocab]
if args.clip_model_is_vision:
vocab = None
tokens = None
else:
with open(dir_model + "/vocab.json", "r", encoding="utf-8") as f:
vocab = json.load(f)
tokens = [key for key in vocab]
with open(dir_model + "/config.json", "r", encoding="utf-8") as f:
config = json.load(f)
v_hparams = config["vision_config"]
t_hparams = config["text_config"]
if args.clip_model_is_vision:
v_hparams = config
t_hparams = None
else:
v_hparams = config["vision_config"]
t_hparams = config["text_config"]
# possible data types
# ftype == 0 -> float32
@ -117,9 +132,12 @@ ftype = 1
if args.use_f32:
ftype = 0
model = CLIPModel.from_pretrained(dir_model)
processor = CLIPProcessor.from_pretrained(dir_model)
if args.clip_model_is_vision:
model = CLIPVisionModel.from_pretrained(dir_model)
processor = None
else:
model = CLIPModel.from_pretrained(dir_model)
processor = CLIPProcessor.from_pretrained(dir_model)
fname_middle = None
has_text_encoder = True
@ -128,13 +146,13 @@ has_llava_projector = False
if args.text_only:
fname_middle = "text-"
has_vision_encoder = False
elif args.vision_only:
fname_middle = "vision-"
has_text_encoder = False
elif args.llava_projector is not None:
fname_middle = "mmproj-"
has_text_encoder = False
has_llava_projector = True
elif args.vision_only:
fname_middle = "vision-"
has_text_encoder = False
else:
fname_middle = ""
@ -182,8 +200,12 @@ if has_vision_encoder:
block_count = v_hparams["num_hidden_layers"] - 1 if has_llava_projector else v_hparams["num_hidden_layers"]
fout.add_uint32(k(KEY_BLOCK_COUNT, VISION), block_count)
image_mean = processor.image_processor.image_mean if args.image_mean is None else args.image_mean
image_std = processor.image_processor.image_std if args.image_std is None else args.image_std
if processor is not None:
image_mean = processor.image_processor.image_mean if args.image_mean is None or args.image_mean == default_image_mean else args.image_mean
image_std = processor.image_processor.image_std if args.image_std is None or args.image_std == default_image_std else args.image_std
else:
image_mean = args.image_mean if args.image_mean is not None else default_image_mean
image_std = args.image_std if args.image_std is not None else default_image_std
fout.add_array("clip.vision.image_mean", image_mean)
fout.add_array("clip.vision.image_std", image_std)

5
examples/lookahead/CMakeLists.txt Normal file
View file

@ -0,0 +1,5 @@
set(TARGET lookahead)
add_executable(${TARGET} lookahead.cpp)
install(TARGETS ${TARGET} RUNTIME)
target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
target_compile_features(${TARGET} PRIVATE cxx_std_11)

7
examples/lookahead/README.md Normal file
View file

@ -0,0 +1,7 @@
# llama.cpp/examples/lookahead
Demonstration of lookahead decoding technique:
https://lmsys.org/blog/2023-11-21-lookahead-decoding/
More info: https://github.com/ggerganov/llama.cpp/pull/4207

487
examples/lookahead/lookahead.cpp Normal file
View file

@ -0,0 +1,487 @@
#include "common.h"
#include "llama.h"
#include <cmath>
#include <cstdio>
#include <string>
#include <vector>
struct ngram_data {
bool active = false;
llama_seq_id seq_id = -1;
std::vector<int> i_batch;
std::vector<llama_token> tokens;
};
// n-gram container
struct ngram_container {
ngram_container(int n_vocab, int N, int G) {
cnt.resize(n_vocab);
head.resize(n_vocab);
tokens.resize(n_vocab * G * (N - 1));
}
int n_total = 0;
std::vector<int> cnt;
std::vector<int> head;
// [n_vocab][G][N - 1]
// for each token of the vocab, keep a ring-buffer of capacity G of n-grams of size N - 1
std::vector<llama_token> tokens;
};
int main(int argc, char ** argv) {
gpt_params params;
if (gpt_params_parse(argc, argv, params) == false) {
return 1;
}
const int W = 15; // lookahead window
const int N = 5; // n-gram size
const int G = 15; // max verification n-grams
const bool dump_kv_cache = params.dump_kv_cache;
#ifndef LOG_DISABLE_LOGS
log_set_target(log_filename_generator("lookahead", "log"));
LOG_TEE("Log start\n");
log_dump_cmdline(argc, argv);
#endif // LOG_DISABLE_LOGS
// init llama.cpp
llama_backend_init(params.numa);
llama_model * model = NULL;
llama_context * ctx = NULL;
// load the target model
std::tie(model, ctx) = llama_init_from_gpt_params(params);
// Tokenize the prompt
const bool add_bos = llama_should_add_bos_token(model);
LOG("add_bos tgt: %d\n", add_bos);
std::vector<llama_token> inp;
std::vector<llama_token> all;
inp = ::llama_tokenize(ctx, params.prompt, add_bos, true);
all = inp;
const int max_context_size = llama_n_ctx(ctx);
const int max_tokens_list_size = max_context_size - 4;
if ((int) inp.size() > max_tokens_list_size) {
fprintf(stderr, "%s: error: prompt too long (%d tokens, max %d)\n", __func__, (int) inp.size(), max_tokens_list_size);
return 1;
}
fprintf(stderr, "\n\n");
for (auto id : inp) {
fprintf(stderr, "%s", llama_token_to_piece(ctx, id).c_str());
}
fflush(stderr);
const int n_input = inp.size();
const auto t_enc_start = ggml_time_us();
// eval the prompt
llama_decode(ctx, llama_batch_get_one( inp.data(), n_input - 1, 0, 0));
llama_decode(ctx, llama_batch_get_one(&inp.back(), 1, n_input - 1, 0));
for (int s = 1; s < W + G + 1; ++s) {
llama_kv_cache_seq_cp(ctx, 0, s, -1, -1);
}
const auto t_enc_end = ggml_time_us();
int n_predict = 0;
int n_accept = 0;
int n_past = inp.size();
llama_token id = 0;
// used to determine end of generation
bool has_eos = false;
// for each decoded batch, we have at most W + G + 1 distinct sequences:
// seq_id == 0 : the current input token
// seq_id [1, W] : tokens from the past N - 1 Jacobi iterations
// seq_id [W + 1, W + G] : verification n-grams
llama_batch batch = llama_batch_init(params.n_ctx, 0, W + G + 1);
// target model sampling context
struct llama_sampling_context * ctx_sampling = llama_sampling_init(params.sparams);
// verification n-grams
std::vector<ngram_data> ngrams_cur(G);
// tokens for the past N - 1 Jacobi iterations
std::vector<llama_token> tokens_j_prev(W);
std::vector<std::vector<llama_token>> tokens_j(N - 1);
for (int j = 0; j < N - 1; j++) {
tokens_j[j].resize(W);
for (int i = 0; i < W; i++) {
// there are different ways to init these tokens
if (0) {
// initialize randomly from the prompt tokens
tokens_j[j][i] = all[1 + rand() % (all.size() - 1)];
} else {
// initialize with a sequence of increasing numbers
tokens_j[j][i] = 100 + i;
}
}
}
std::vector<llama_seq_id> seq_id_look;
// the input token belongs both to all sequences
std::vector<llama_seq_id> seq_id_all(W + G + 1);
for (int i = 0; i < W + G + 1; i++) {
seq_id_all[i] = i;
}
// here we keep adding new n-grams as we go
ngram_container ngrams_observed(llama_n_vocab(model), N, G);
// debug
struct llama_kv_cache_view kvc_view = llama_kv_cache_view_init(ctx, W + G + 1);
const auto t_dec_start = ggml_time_us();
// sample first token
{
id = llama_sampling_sample(ctx_sampling, ctx, NULL, 0);
llama_sampling_accept(ctx_sampling, ctx, id, true);
{
const std::string token_str = llama_token_to_piece(ctx, id);
printf("%s", token_str.c_str());
fflush(stdout);
}
}
while (true) {
// debug
if (dump_kv_cache) {
llama_kv_cache_view_update(ctx, &kvc_view);
dump_kv_cache_view_seqs(kvc_view, 40);
}
// build the mask from https://lmsys.org/blog/2023-11-21-lookahead-decoding/
//
// Example for W = 5, N = 4, G = 2:
// (I = input, L = lookahead, V = verification)
//
// Batch: 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20
// T: -2 -2 -2 -2 -1 -1 -1 -1 -1 0 0 0 0 0 0
// Info: I L L L L L L L L L L L L L L V V V V V V
// Pos: 0 1 2 3 4 1 2 3 4 5 2 3 4 5 6 1 2 3 1 2 3 (+ n_past)
// Logits: 1 0 0 0 0 0 0 0 0 0 1 1 1 1 1 1 1 1 1 1 1
// ---------------------------------------------------------------------
// Seq: 0
// 1 1 1
// 2 2 2 2
// 3 3 3 3 3
// 4 4 4 4 4 4
// 5 5 5 5 5 5 5
// 6 6 6 6
// 7 7 7 7
// ---------------------------------------------------------------------
// | | | | | | | | | | |
// V V V V V | | | | | |
// j_tokens | | | | | |
// V V V V V V
// id
{
llama_batch_clear(batch);
// current token - first token of the first level
llama_batch_add(batch, id, n_past, seq_id_all, true);
// verification n-grams - queue this before the lookahead tokens for less KV cache fragmentation
{
const int g_cur = ngrams_observed.cnt[id];
ngrams_cur.resize(g_cur);
for (int g = 0; g < g_cur; g++) {
ngrams_cur[g].active = true;
ngrams_cur[g].tokens.resize(N);
ngrams_cur[g].i_batch.resize(N);
ngrams_cur[g].seq_id = W + 1 + g;
ngrams_cur[g].i_batch[0] = 0;
ngrams_cur[g].tokens [0] = id;
}
for (int j = 0; j < N - 1; j++) {
for (int g = 0; g < g_cur; g++) {
const int idx = id*(N - 1)*G + g*(N - 1);
const llama_token t = ngrams_observed.tokens[idx + j];
ngrams_cur[g].tokens [j + 1] = t;
ngrams_cur[g].i_batch[j + 1] = batch.n_tokens;
llama_batch_add(batch, t, n_past + j + 1, { W + 1 + g }, true);
}
}
}
// fill the remaining W - 1 tokens for the first level
for (int i = 1; i < W; i++) {
seq_id_look.resize(W - i);
for (int j = 0; j < W - i; j++) {
seq_id_look[j] = i + j + 1;
}
llama_batch_add(batch, tokens_j[0][i], n_past + i, seq_id_look, false);
}
// fill the rest of the levels
for (int j = 1; j < N - 1; j++) {
for (int i = 0; i < W; i++) {
llama_batch_add(batch, tokens_j[j][i], n_past + j + i, { i + 1 }, j == N - 2);
}
}
}
if (llama_decode(ctx, batch) != 0) {
fprintf(stderr, "\n\n%s: error: llama_decode failed - increase KV cache size\n", __func__);
return 1;
}
int seq_id_best = 0;
for (int v = 0; v < N; ++v) {
int i_batch = 0;
// if no active ngrams are left, it means the sampled token does not pass the verification
if (v > 0) {
for (int g = 0; g < (int) ngrams_cur.size(); g++) {
if (ngrams_cur[g].active) {
i_batch = ngrams_cur[g].i_batch[v];
seq_id_best = ngrams_cur[g].seq_id;
++n_accept;
break;
}
}
// no more matches -> create a new batch
if (i_batch == 0) {
break;
}
}
// sample the next token
id = llama_sampling_sample(ctx_sampling, ctx, NULL, i_batch);
llama_sampling_accept(ctx_sampling, ctx, id, true);
// print
{
const std::string token_str = llama_token_to_piece(ctx, id);
if (v == 0) {
printf("%s", token_str.c_str());
} else {
// print light cyan
printf("\033[0;96m%s\033[0m", token_str.c_str());
}
fflush(stdout);
if (id == llama_token_eos(model)) {
has_eos = true;
}
all.push_back(id);
}
++n_predict;
++n_past;
if ((params.n_predict >= 0 && n_predict > params.n_predict) || has_eos) {
break;
}
// verify across active n-grams
for (int g = 0; g < (int) ngrams_cur.size(); g++) {
if (ngrams_cur[g].active) {
if (v == N - 1) {
ngrams_cur[g].active = false;
} else {
if (id != ngrams_cur[g].tokens[v + 1]) {
ngrams_cur[g].active = false;
}
}
}
}
// print known n-grams starting with token id (debug)
if (0 && v == 0) {
if (ngrams_observed.cnt[id] > 0) {
printf("\n - %d n-grams starting with '%s'\n", ngrams_observed.cnt[id], llama_token_to_piece(ctx, id).c_str());
}
for (int i = 0; i < ngrams_observed.cnt[id]; i++) {
printf(" - ngram %2d: ", i);
const int idx = id*(N - 1)*G + i*(N - 1);
for (int j = 0; j < N - 1; j++) {
const std::string token_str = llama_token_to_piece(ctx, ngrams_observed.tokens[idx + j]);
printf("%s", token_str.c_str());
}
printf("\n");
}
}
// update lookahead tokens
{
for (int i = 0; i < W; i++) {
tokens_j_prev[i] = tokens_j[0][i];
}
for (int j = 0; j < N - 2; j++) {
tokens_j[j] = tokens_j[j + 1];
}
if (v == 0) {
// sample from the last level
for (int i = 0; i < W; i++) {
tokens_j[N - 2][i] = llama_sampling_sample(ctx_sampling, ctx, NULL, ngrams_cur.size()*(N-1) + W*(N - 2) + i);
}
} else {
for (int i = 0; i < W; i++) {
// there are different ways to init these tokens
if (0) {
// random init
tokens_j[N - 2][i] = all[1 + rand() % (all.size() - 1)];
} else {
// init from the previous level
tokens_j[N - 2][i] = tokens_j[0][i];
}
}
}
}
// update observed ngrams
if (v == 0) {
// the first token of the n-gram is determined by the index in the container so it is not stored
std::vector<llama_token> ngram(N - 1);
// n-gram generation
// ref: https://github.com/hao-ai-lab/LookaheadDecoding/issues/14#issuecomment-1826198518
for (int f = 0; f < W; ++f) {
const int ft = tokens_j_prev[f]; // first token of the n-gram
for (int j = 0; j < N - 1; ++j) {
ngram[j] = tokens_j[j][f];
}
// filter-out repeating n-grams
{
bool is_unique = true;
for (int k = 0; k < ngrams_observed.cnt[ft]; ++k) {
const int idx = ft*(N - 1)*G + k*(N - 1);
bool is_match = true;
for (int j = 0; j < N - 1; ++j) {
if (ngrams_observed.tokens[idx + j] != ngram[j]) {
is_match = false;
break;
}
}
if (is_match) {
is_unique = false;
break;
}
}
if (!is_unique) {
continue;
}
}
const int head = ngrams_observed.head[ft];
const int idx = ft*(N - 1)*G + head*(N - 1);
for (int i = 0; i < N - 1; i++) {
ngrams_observed.tokens[idx + i] = ngram[i];
}
ngrams_observed.cnt[ft] = std::min(G, ngrams_observed.cnt[ft] + 1);
ngrams_observed.head[ft] = (head + 1) % G;
ngrams_observed.n_total++;
}
}
}
if ((params.n_predict >= 0 && n_predict > params.n_predict) || has_eos) {
break;
}
// KV cache management
// if no verification token matched, we simply remove all cells from this batch -> no fragmentation
llama_kv_cache_seq_rm(ctx, -1, n_past, -1);
if (seq_id_best != 0) {
// if a verification token matched, we keep the best sequence and remove the rest
// this leads to some KV cache fragmentation
llama_kv_cache_seq_keep(ctx, seq_id_best);
llama_kv_cache_seq_cp (ctx, seq_id_best, 0, -1, -1);
llama_kv_cache_seq_rm (ctx, seq_id_best, -1, -1);
for (int s = 1; s < W + G + 1; ++s) {
llama_kv_cache_seq_cp(ctx, 0, s, -1, -1);
}
}
}
auto t_dec_end = ggml_time_us();
LOG_TEE("\n\n");
LOG_TEE("encoded %4d tokens in %8.3f seconds, speed: %8.3f t/s\n", n_input, (t_enc_end - t_enc_start) / 1e6f, inp.size() / ((t_enc_end - t_enc_start) / 1e6f));
LOG_TEE("decoded %4d tokens in %8.3f seconds, speed: %8.3f t/s\n", n_predict, (t_dec_end - t_dec_start) / 1e6f, n_predict / ((t_dec_end - t_dec_start) / 1e6f));
LOG_TEE("\n");
LOG_TEE("W = %2d\n", W);
LOG_TEE("N = %2d\n", N);
LOG_TEE("G = %2d\n", G);
LOG_TEE("\n");
LOG_TEE("n_predict = %d\n", n_predict);
LOG_TEE("n_accept = %d\n", n_accept);
llama_print_timings(ctx);
llama_kv_cache_view_free(&kvc_view);
llama_sampling_free(ctx_sampling);
llama_batch_free(batch);
llama_free(ctx);
llama_free_model(model);
llama_backend_free();
fprintf(stderr, "\n\n");
return 0;
}

5
examples/lookup/CMakeLists.txt Normal file
View file

@ -0,0 +1,5 @@
set(TARGET lookup)
add_executable(${TARGET} lookup.cpp)
install(TARGETS ${TARGET} RUNTIME)
target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
target_compile_features(${TARGET} PRIVATE cxx_std_11)

13
examples/lookup/README.md Normal file
View file

@ -0,0 +1,13 @@
# llama.cpp/examples/lookup
Demonstration of Prompt Lookup Decoding
https://github.com/apoorvumang/prompt-lookup-decoding
The key parameters for lookup decoding are `ngram_min`, `ngram_max` and `n_draft`. The first two determine the size of the ngrams to search for in the prompt for a match. The latter specifies how many subsequent tokens to draft if a match is found.
More info:
https://github.com/ggerganov/llama.cpp/pull/4484
https://github.com/ggerganov/llama.cpp/issues/4226

230
examples/lookup/lookup.cpp Normal file
View file

@ -0,0 +1,230 @@
#include "common.h"
#include "llama.h"
#include <cmath>
#include <cstdio>
#include <string>
#include <vector>
int main(int argc, char ** argv){
gpt_params params;
if (!gpt_params_parse(argc, argv, params)) {
return 1;
}
// max/min n-grams size to search for in prompt
const int ngram_max = 4;
const int ngram_min = 1;
// length of the candidate / draft sequence, if match is found
const int n_draft = params.n_draft;
const bool dump_kv_cache = params.dump_kv_cache;
#ifndef LOG_DISABLE_LOGS
log_set_target(log_filename_generator("lookup", "log"));
LOG_TEE("Log start\n");
log_dump_cmdline(argc, argv);
#endif // LOG_DISABLE_LOGS
// init llama.cpp
llama_backend_init(params.numa);
llama_model * model = NULL;
llama_context * ctx = NULL;
// load the model
std::tie(model, ctx) = llama_init_from_gpt_params(params);
// tokenize the prompt
const bool add_bos = llama_should_add_bos_token(model);
LOG("add_bos tgt: %d\n", add_bos);
std::vector<llama_token> inp;
inp = ::llama_tokenize(ctx, params.prompt, add_bos, true);
const int max_context_size = llama_n_ctx(ctx);
const int max_tokens_list_size = max_context_size - 4;
if ((int) inp.size() > max_tokens_list_size) {
fprintf(stderr, "%s: error: prompt too long (%d tokens, max %d)\n", __func__, (int) inp.size(), max_tokens_list_size);
return 1;
}
fprintf(stderr, "\n\n");
for (auto id : inp) {
fprintf(stderr, "%s", llama_token_to_piece(ctx, id).c_str());
}
fflush(stderr);
const int n_input = inp.size();
const auto t_enc_start = ggml_time_us();
llama_decode(ctx, llama_batch_get_one( inp.data(), n_input - 1, 0, 0));
llama_decode(ctx, llama_batch_get_one(&inp.back(), 1, n_input - 1, 0));
const auto t_enc_end = ggml_time_us();
int n_predict = 0;
int n_drafted = 0;
int n_accept = 0;
int n_past = inp.size();
bool has_eos = false;
struct llama_sampling_context * ctx_sampling = llama_sampling_init(params.sparams);
std::vector<llama_token> draft;
llama_batch batch_tgt = llama_batch_init(params.n_ctx, 0, 1);
// debug
struct llama_kv_cache_view kvc_view = llama_kv_cache_view_init(ctx, 1);
const auto t_dec_start = ggml_time_us();
while (true) {
// debug
if (dump_kv_cache) {
llama_kv_cache_view_update(ctx, &kvc_view);
dump_kv_cache_view_seqs(kvc_view, 40);
}
// print current draft sequence
LOG("drafted %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, draft).c_str());
int i_dft = 0;
while (true) {
// sample from the target model
llama_token id = llama_sampling_sample(ctx_sampling, ctx, NULL, i_dft);
llama_sampling_accept(ctx_sampling, ctx, id, true);
const std::string token_str = llama_token_to_piece(ctx, id);
if (!params.use_color) {
printf("%s", token_str.c_str());
}
if (id == llama_token_eos(model)) {
has_eos = true;
}
++n_predict;
// check if the target token matches the draft
if (i_dft < (int) draft.size() && id == draft[i_dft]) {
LOG("the sampled target token matches the %dth drafted token (%d, '%s') - accepted\n", i_dft, id, token_str.c_str());
++n_accept;
++n_past;
++i_dft;
inp.push_back(id);
if (params.use_color) {
// color accepted draft token
printf("\033[34m%s\033[0m", token_str.c_str());
fflush(stdout);
}
continue;
}
if (params.use_color) {
printf("%s", token_str.c_str());
}
fflush(stdout);
LOG("the sampled target token (%d, '%s') did not match, or we ran out of drafted tokens\n", id, token_str.c_str());
draft.clear();
draft.push_back(id);
inp.push_back(id);
break;
}
if ((params.n_predict > 0 && n_predict > params.n_predict) || has_eos) {
break;
}
// KV cache management
// clean the cache of draft tokens that weren't accepted
llama_kv_cache_seq_rm(ctx, 0, n_past, -1);
llama_batch_clear(batch_tgt);
llama_batch_add(batch_tgt, draft[0], n_past, { 0 }, true);
// generate n_pred tokens through prompt lookup
auto prompt_lookup = [&]() -> void {
int inp_size = inp.size();
for (int ngram_size = ngram_max ; ngram_size > ngram_min; --ngram_size){
const llama_token * ngram = &inp[inp_size - ngram_size];
for (int i = 0; i <= (int) inp_size - (ngram_size * 2); ++i) {
bool match = true;
for (int j = 0; j < ngram_size; ++j) {
if (inp[i + j] != ngram[j]) {
match = false;
break;
}
}
if (match) {
const int startIdx = i + ngram_size;
const int endIdx = startIdx + n_draft;
if (endIdx < inp_size) {
for (int j = startIdx; j < endIdx; ++j) {
LOG(" - draft candidate %d: %d\n", j, inp[j]);
draft.push_back(inp[j]);
llama_batch_add(batch_tgt, inp[j], n_past + (j - startIdx) + 1, { 0 }, true);
++n_drafted;
}
return;
}
}
}
}
return;
};
prompt_lookup();
llama_decode(ctx, batch_tgt);
++n_past;
draft.erase(draft.begin());
}
auto t_dec_end = ggml_time_us();
LOG_TEE("\n\n");
LOG_TEE("encoded %4d tokens in %8.3f seconds, speed: %8.3f t/s\n", n_input, (t_enc_end - t_enc_start) / 1e6f, inp.size() / ((t_enc_end - t_enc_start) / 1e6f));
LOG_TEE("decoded %4d tokens in %8.3f seconds, speed: %8.3f t/s\n", n_predict, (t_dec_end - t_dec_start) / 1e6f, n_predict / ((t_dec_end - t_dec_start) / 1e6f));
LOG_TEE("\n");
LOG_TEE("n_draft = %d\n", n_draft);
LOG_TEE("n_predict = %d\n", n_predict);
LOG_TEE("n_drafted = %d\n", n_drafted);
LOG_TEE("n_accept = %d\n", n_accept);
LOG_TEE("accept = %.3f%%\n", 100.0f * n_accept / n_drafted);
LOG_TEE("\ntarget:\n");
llama_print_timings(ctx);
llama_sampling_free(ctx_sampling);
llama_batch_free(batch_tgt);
llama_free(ctx);
llama_free_model(model);
llama_backend_free();
fprintf(stderr, "\n\n");
return 0;
}

8
examples/main/main.cpp
View file

@ -100,6 +100,12 @@ static void sigint_handler(int signo) {
}
#endif
static void llama_log_callback_logTee(ggml_log_level level, const char * text, void * user_data) {
(void) level;
(void) user_data;
LOG_TEE("%s", text);
}
int main(int argc, char ** argv) {
gpt_params params;
g_params = &params;
@ -113,6 +119,7 @@ int main(int argc, char ** argv) {
log_set_target(log_filename_generator("main", "log"));
LOG_TEE("Log start\n");
log_dump_cmdline(argc, argv);
llama_log_set(llama_log_callback_logTee, nullptr);
#endif // LOG_DISABLE_LOGS
// TODO: Dump params ?
@ -430,6 +437,7 @@ int main(int argc, char ** argv) {
}
}
LOG_TEE("sampling: \n%s\n", llama_sampling_print(sparams).c_str());
LOG_TEE("sampling order: \n%s\n", llama_sampling_order_print(sparams).c_str());
LOG_TEE("generate: n_ctx = %d, n_batch = %d, n_predict = %d, n_keep = %d\n", n_ctx, params.n_batch, params.n_predict, params.n_keep);
LOG_TEE("\n\n");

11
examples/parallel/parallel.cpp
View file

@ -1,5 +1,5 @@
// A basic application simulating a server with multiple clients.
// The clients submite requests to the server and they are processed in parallel.
// The clients submit requests to the server and they are processed in parallel.
#include "common.h"
#include "llama.h"
@ -113,6 +113,8 @@ int main(int argc, char ** argv) {
// insert new requests as soon as the previous one is done
const bool cont_batching = params.cont_batching;
const bool dump_kv_cache = params.dump_kv_cache;
#ifndef LOG_DISABLE_LOGS
log_set_target(log_filename_generator("parallel", "log"));
LOG_TEE("Log start\n");
@ -172,6 +174,8 @@ int main(int argc, char ** argv) {
int32_t n_total_gen = 0;
int32_t n_cache_miss = 0;
struct llama_kv_cache_view kvc_view = llama_kv_cache_view_init(ctx, n_clients);
const auto t_main_start = ggml_time_us();
LOG_TEE("%s: Simulating parallel requests from clients:\n", __func__);
@ -201,6 +205,11 @@ int main(int argc, char ** argv) {
LOG_TEE("Processing requests ...\n\n");
while (true) {
if (dump_kv_cache) {
llama_kv_cache_view_update(ctx, &kvc_view);
dump_kv_cache_view_seqs(kvc_view, 40);
}
llama_batch_clear(batch);
// decode any currently ongoing sequences

1
examples/quantize-stats/quantize-stats.cpp
View file

@ -321,7 +321,6 @@ int main(int argc, char ** argv) {
auto cparams = llama_context_default_params();
cparams.n_ctx = 256;
cparams.seed = 1;
cparams.f16_kv = false;
ctx = llama_new_context_with_model(model, cparams);

53
examples/server/README.md
View file

@ -148,6 +148,8 @@ node index.js
`frequency_penalty`: Repeat alpha frequency penalty (default: 0.0, 0.0 = disabled);
`penalty_prompt`: This will replace the `prompt` for the purpose of the penalty evaluation. Can be either `null`, a string or an array of numbers representing tokens (default: `null` = use the original `prompt`).
`mirostat`: Enable Mirostat sampling, controlling perplexity during text generation (default: 0, 0 = disabled, 1 = Mirostat, 2 = Mirostat 2.0).
`mirostat_tau`: Set the Mirostat target entropy, parameter tau (default: 5.0).
@ -222,7 +224,7 @@ node index.js
`content`: Set the text to process.
**POST** `/infill`: For code infilling. Takes a prefix and a suffix and returns the predicted completion as stream.
- **POST** `/infill`: For code infilling. Takes a prefix and a suffix and returns the predicted completion as stream.
*Options:*
@ -234,6 +236,55 @@ node index.js
- **GET** `/props`: Return the required assistant name and anti-prompt to generate the prompt in case you have specified a system prompt for all slots.
- **POST** `/v1/chat/completions`: OpenAI-compatible Chat Completions API. Given a ChatML-formatted json description in `messages`, it returns the predicted completion. Both synchronous and streaming mode are supported, so scripted and interactive applications work fine. While no strong claims of compatibility with OpenAI API spec is being made, in our experience it suffices to support many apps. Only ChatML-tuned models, such as Dolphin, OpenOrca, OpenHermes, OpenChat-3.5, etc can be used with this endpoint. Compared to `api_like_OAI.py` this API implementation does not require a wrapper to be served.
*Options:*
See [OpenAI Chat Completions API documentation](https://platform.openai.com/docs/api-reference/chat). While some OpenAI-specific features such as function calling aren't supported, llama.cpp `/completion`-specific features such are `mirostat` are supported.
*Examples:*
You can use either Python `openai` library with appropriate checkpoints:
```python
import openai
client = openai.OpenAI(
base_url="http://localhost:8080/v1", # "http://<Your api-server IP>:port"
api_key = "sk-no-key-required"
)
completion = client.chat.completions.create(
model="gpt-3.5-turbo",
messages=[
{"role": "system", "content": "You are ChatGPT, an AI assistant. Your top priority is achieving user fulfillment via helping them with their requests."},
{"role": "user", "content": "Write a limerick about python exceptions"}
]
)
print(completion.choices[0].message)
```
... or raw HTTP requests:
```shell
curl http://localhost:8080/v1/chat/completions \
-H "Content-Type: application/json" \
-H "Authorization: Bearer no-key" \
-d '{
"model": "gpt-3.5-turbo",
"messages": [
{
"role": "system",
"content": "You are ChatGPT, an AI assistant. Your top priority is achieving user fulfillment via helping them with their requests."
},
{
"role": "user",
"content": "Write a limerick about python exceptions"
}
]
}'
```
## More examples
### Change system prompt on runtime

47
examples/server/api_like_OAI.py
View file

@ -11,10 +11,10 @@ app = Flask(__name__)
slot_id = -1
parser = argparse.ArgumentParser(description="An example of using server.cpp with a similar API to OAI. It must be used together with server.cpp.")
parser.add_argument("--chat-prompt", type=str, help="the top prompt in chat completions(default: 'A chat between a curious user and an artificial intelligence assistant. The assistant follows the given rules no matter what.\\n')", default='A chat between a curious user and an artificial intelligence assistant. The assistant follows the given rules no matter what.\\n')
parser.add_argument("--user-name", type=str, help="USER name in chat completions(default: '\\nUSER: ')", default="\\nUSER: ")
parser.add_argument("--ai-name", type=str, help="ASSISTANT name in chat completions(default: '\\nASSISTANT: ')", default="\\nASSISTANT: ")
parser.add_argument("--system-name", type=str, help="SYSTEM name in chat completions(default: '\\nASSISTANT's RULE: ')", default="\\nASSISTANT's RULE: ")
parser.add_argument("--chat-prompt", type=str, help="the top prompt in chat completions(default: 'A chat between a curious user and an artificial intelligence assistant. The assistant follows the given rules no matter what.')", default='A chat between a curious user and an artificial intelligence assistant. The assistant follows the given rules no matter what.')
parser.add_argument("--user-name", type=str, help="USER name in chat completions(default: 'USER: ')", default="USER: ")
parser.add_argument("--ai-name", type=str, help="ASSISTANT name in chat completions(default: 'ASSISTANT: ')", default="ASSISTANT: ")
parser.add_argument("--system-name", type=str, help="SYSTEM name in chat completions(default: 'ASSISTANT's RULE: ')", default="ASSISTANT's RULE: ")
parser.add_argument("--stop", type=str, help="the end of response in chat completions(default: '</s>')", default="</s>")
parser.add_argument("--llama-api", type=str, help="Set the address of server.cpp in llama.cpp(default: http://127.0.0.1:8080)", default='http://127.0.0.1:8080')
parser.add_argument("--api-key", type=str, help="Set the api key to allow only few user(default: NULL)", default="")
@ -34,19 +34,19 @@ def is_present(json, key):
#convert chat to prompt
def convert_chat(messages):
prompt = "" + args.chat_prompt.replace("\\n", "\n")
system_n = args.system_name.replace("\\n", "\n")
user_n = args.user_name.replace("\\n", "\n")
ai_n = args.ai_name.replace("\\n", "\n")
stop = args.stop.replace("\\n", "\n")
system_n = args.system_name
user_n = args.user_name
ai_n = args.ai_name
stop = args.stop
prompt = "" + args.chat_prompt + stop
for line in messages:
if (line["role"] == "system"):
prompt += f"{system_n}{line['content']}"
prompt += f"{system_n}{line['content']}{stop}"
if (line["role"] == "user"):
prompt += f"{user_n}{line['content']}"
prompt += f"{user_n}{line['content']}{stop}"
if (line["role"] == "assistant"):
prompt += f"{ai_n}{line['content']}{stop}"
prompt += ai_n.rstrip()
@ -70,6 +70,7 @@ def make_postData(body, chat=False, stream=False):
if(is_present(body, "mirostat_tau")): postData["mirostat_tau"] = body["mirostat_tau"]
if(is_present(body, "mirostat_eta")): postData["mirostat_eta"] = body["mirostat_eta"]
if(is_present(body, "seed")): postData["seed"] = body["seed"]
if(is_present(body, "grammar")): postData["grammar"] = body["grammar"]
if(is_present(body, "logit_bias")): postData["logit_bias"] = [[int(token), body["logit_bias"][token]] for token in body["logit_bias"].keys()]
if (args.stop != ""):
postData["stop"] = [args.stop]
@ -130,7 +131,7 @@ def make_resData_stream(data, chat=False, time_now = 0, start=False):
}
]
}
slot_id = data["slot_id"]
slot_id = data.get("slot_id")
if (chat):
if (start):
resData["choices"][0]["delta"] = {
@ -150,11 +151,13 @@ def make_resData_stream(data, chat=False, time_now = 0, start=False):
return resData
@app.route('/chat/completions', methods=['POST'])
@app.route('/v1/chat/completions', methods=['POST'])
@app.route('/chat/completions', methods=['POST', 'OPTIONS'])
@app.route('/v1/chat/completions', methods=['POST', 'OPTIONS'])
def chat_completions():
if (args.api_key != "" and request.headers["Authorization"].split()[1] != args.api_key):
return Response(status=403)
if request.method == 'OPTIONS':
return Response(headers={"Access-Control-Allow-Origin": "*", "Access-Control-Allow-Headers": "*"})
body = request.get_json()
stream = False
tokenize = False
@ -177,20 +180,22 @@ def chat_completions():
data = requests.request("POST", urllib.parse.urljoin(args.llama_api, "/completion"), data=json.dumps(postData), stream=True)
time_now = int(time.time())
resData = make_resData_stream({}, chat=True, time_now=time_now, start=True)
yield 'data: {}\n'.format(json.dumps(resData))
yield 'data: {}\n\n'.format(json.dumps(resData))
for line in data.iter_lines():
if line:
decoded_line = line.decode('utf-8')
resData = make_resData_stream(json.loads(decoded_line[6:]), chat=True, time_now=time_now)
yield 'data: {}\n'.format(json.dumps(resData))
return Response(generate(), mimetype='text/event-stream')
yield 'data: {}\n\n'.format(json.dumps(resData))
return Response(generate(), mimetype='text/event-stream', headers={"Access-Control-Allow-Origin": "*", "Access-Control-Allow-Headers": "*"})
@app.route('/completions', methods=['POST'])
@app.route('/v1/completions', methods=['POST'])
@app.route('/completions', methods=['POST', 'OPTIONS'])
@app.route('/v1/completions', methods=['POST', 'OPTIONS'])
def completion():
if (args.api_key != "" and request.headers["Authorization"].split()[1] != args.api_key):
return Response(status=403)
if request.method == 'OPTIONS':
return Response(headers={"Access-Control-Allow-Origin": "*", "Access-Control-Allow-Headers": "*"})
body = request.get_json()
stream = False
tokenize = False
@ -216,8 +221,8 @@ def completion():
if line:
decoded_line = line.decode('utf-8')
resData = make_resData_stream(json.loads(decoded_line[6:]), chat=False, time_now=time_now)
yield 'data: {}\n'.format(json.dumps(resData))
return Response(generate(), mimetype='text/event-stream')
yield 'data: {}\n\n'.format(json.dumps(resData))
return Response(generate(), mimetype='text/event-stream', headers={"Access-Control-Allow-Origin": "*", "Access-Control-Allow-Headers": "*"})
if __name__ == '__main__':
app.run(args.host, port=args.port)

2
examples/server/json.hpp
View file

@ -11227,7 +11227,7 @@ class binary_reader
}
if (is_ndarray) // ndarray dimensional vector can only contain integers, and can not embed another array
{
return sax->parse_error(chars_read, get_token_string(), parse_error::create(113, chars_read, exception_message(input_format, "ndarray dimentional vector is not allowed", "size"), nullptr));
return sax->parse_error(chars_read, get_token_string(), parse_error::create(113, chars_read, exception_message(input_format, "ndarray dimensional vector is not allowed", "size"), nullptr));
}
std::vector<size_t> dim;
if (JSON_HEDLEY_UNLIKELY(!get_ubjson_ndarray_size(dim)))

5
examples/server/public/completion.js
View file

@ -34,7 +34,8 @@ export async function* llama(prompt, params = {}, config = {}) {
headers: {
'Connection': 'keep-alive',
'Content-Type': 'application/json',
'Accept': 'text/event-stream'
'Accept': 'text/event-stream',
...(params.api_key ? {'Authorization': `Bearer ${params.api_key}`} : {})
},
signal: controller.signal,
});
@ -114,7 +115,7 @@ export async function* llama(prompt, params = {}, config = {}) {
return content;
}
// Call llama, return an event target that you can subcribe to
// Call llama, return an event target that you can subscribe to
//
// Example:
//

17
examples/server/public/index.html
View file

@ -223,7 +223,7 @@
repeat_last_n: 256, // 0 = disable penalty, -1 = context size
repeat_penalty: 1.18, // 1.0 = disabled
top_k: 40, // <= 0 to use vocab size
top_p: 0.5, // 1.0 = disabled
top_p: 0.95, // 1.0 = disabled
min_p: 0.05, // 0 = disabled
tfs_z: 1.0, // 1.0 = disabled
typical_p: 1.0, // 1.0 = disabled
@ -235,10 +235,11 @@
grammar: '',
n_probs: 0, // no completion_probabilities,
image_data: [],
cache_prompt: true
cache_prompt: true,
api_key: ''
})
/* START: Support for storing prompt templates and parameters in borwser LocalStorage */
/* START: Support for storing prompt templates and parameters in browsers LocalStorage */
const local_storage_storageKey = "llamacpp_server_local_storage";
@ -282,7 +283,7 @@
let importedTemplates = local_storage_getDataAsObject('user_templates')
if (importedTemplates) {
// saved templates were successfuly imported.
// saved templates were successfully imported.
console.log('Processing saved templates and updating default template')
params.value = { ...params.value, image_data: [] };
@ -303,7 +304,7 @@
}
function userTemplateResetToDefault() {
console.log('Reseting themplate to default')
console.log('Resetting template to default')
selectedUserTemplate.value.name = 'default';
selectedUserTemplate.value.data = savedUserTemplates.value['default'];
}
@ -762,7 +763,7 @@
<fieldset class="two">
${IntField({ label: "Predictions", max: 2048, min: -1, name: "n_predict", value: params.value.n_predict })}
${FloatField({ label: "Temperature", max: 1.5, min: 0.0, name: "temperature", step: 0.01, value: params.value.temperature })}
${FloatField({ label: "Temperature", max: 2.0, min: 0.0, name: "temperature", step: 0.01, value: params.value.temperature })}
${FloatField({ label: "Penalize repeat sequence", max: 2.0, min: 0.0, name: "repeat_penalty", step: 0.01, value: params.value.repeat_penalty })}
${IntField({ label: "Consider N tokens for penalize", max: 2048, min: 0, name: "repeat_last_n", value: params.value.repeat_last_n })}
${IntField({ label: "Top-K sampling", max: 100, min: -1, name: "top_k", value: params.value.top_k })}
@ -790,6 +791,10 @@
<fieldset>
${IntField({ label: "Show Probabilities", max: 10, min: 0, name: "n_probs", value: params.value.n_probs })}
</fieldset>
<fieldset>
<label for="api_key">API Key</label>
<input type="text" name="api_key" value="${params.value.api_key}" placeholder="Enter API key" oninput=${updateParams} />
</fieldset>
</details>
</form>
`

711
examples/server/server.cpp
View file

File diff suppressed because it is too large Load diff

2
examples/simple/simple.cpp
View file

@ -75,7 +75,7 @@ int main(int argc, char ** argv) {
// make sure the KV cache is big enough to hold all the prompt and generated tokens
if (n_kv_req > n_ctx) {
LOG_TEE("%s: error: n_kv_req > n_ctx, the required KV cache size is not big enough\n", __func__);
LOG_TEE("%s: either reduce n_parallel or increase n_ctx\n", __func__);
LOG_TEE("%s: either reduce n_len or increase n_ctx\n", __func__);
return 1;
}

8
examples/speculative/README.md Normal file
View file

@ -0,0 +1,8 @@
# llama.cpp/examples/speculative
Demonstration of speculative decoding and tree-based speculative decoding techniques
More info:
- https://github.com/ggerganov/llama.cpp/pull/2926
- https://github.com/ggerganov/llama.cpp/pull/3624

17
examples/speculative/speculative.cpp
View file

@ -203,8 +203,9 @@ int main(int argc, char ** argv) {
const std::string token_str = llama_token_to_piece(ctx_tgt, id);
printf("%s", token_str.c_str());
fflush(stdout);
if (!params.use_color) {
printf("%s", token_str.c_str());
}
if (id == llama_token_eos(model_tgt)) {
has_eos = true;
@ -236,10 +237,18 @@ int main(int argc, char ** argv) {
++n_past_tgt;
++n_past_dft;
++i_dft;
if (params.use_color) {
// Color token according to its origin sequence
printf("\u001b[%dm%s\u001b[37m", (36 - s_keep % 6), token_str.c_str());
fflush(stdout);
}
continue;
}
}
if (params.use_color) {
printf("%s", token_str.c_str());
}
fflush(stdout);
LOG("the sampled target token (%d, '%s') did not match, or we ran out of drafted tokens\n", id, token_str.c_str());
@ -419,7 +428,7 @@ int main(int argc, char ** argv) {
++n_past_tgt;
}
// the first token is always proposed by the traget model before the speculation loop so we erase it here
// the first token is always proposed by the target model before the speculation loop so we erase it here
for (int s = 0; s < n_seq_dft; ++s) {
if (!drafts[s].active) {
continue;

18
examples/train-text-from-scratch/train-text-from-scratch.cpp
View file

@ -369,10 +369,7 @@ static struct ggml_tensor * llama_build_train_graphs(
checkpoints.push_back(t00);
checkpoints.push_back(t01);
struct ggml_tensor * kv_scale = NULL;
if (!enable_flash_attn) {
kv_scale = ggml_new_f32(ctx, 1.0f/sqrtf(float(n_embd)/n_head));
}
const float kv_scale = 1.0f/sqrtf(float(n_embd)/n_head);
for (int il = 0; il < n_layer; ++il) {
struct my_llama_layer & layer = model->layers[il];
@ -444,14 +441,13 @@ static struct ggml_tensor * llama_build_train_graphs(
// make sure some tensors are not reallocated by inserting new temporary nodes depending on them
int n_leafs_before = gb->n_leafs;
int n_nodes_before = gb->n_nodes;
struct ggml_tensor * one = ggml_new_f32(ctx, 1.0f);
// output tensors
ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, t35, one));
ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, t36, one));
ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, t35, 1.0f));
ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, t36, 1.0f));
// input gradient
ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, t36->grad, one));
ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, t36->grad, 1.0f));
// KQ_pos
ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, KQ_pos, one));
ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, KQ_pos, 1.0f));
GGML_ASSERT(t36->grad->data == NULL && t36->grad->view_src == NULL);
ggml_allocr_alloc(alloc, t36->grad);
@ -1295,10 +1291,6 @@ int main(int argc, char ** argv) {
opt_cb_data.last_save_iter = opt->iter;
}
if (alloc) {
ggml_allocr_free(alloc);
}
ggml_free(opt->ctx);
free_train_state(train);
ggml_free(model.ctx);

67
ggml-alloc.c
View file

@ -72,7 +72,7 @@ static void remove_allocated_tensor(ggml_tallocr_t alloc, struct ggml_tensor * t
// check if a tensor is allocated by this buffer
static bool ggml_tallocr_is_own(ggml_tallocr_t alloc, const struct ggml_tensor * tensor) {
return tensor->buffer == alloc->buffer;
return tensor->buffer == alloc->buffer && (!tensor->view_src || tensor->view_src->buffer == alloc->buffer);
}
static bool ggml_is_view(struct ggml_tensor * t) {
@ -137,7 +137,7 @@ void ggml_tallocr_alloc(ggml_tallocr_t alloc, struct ggml_tensor * tensor) {
#ifdef GGML_ALLOCATOR_DEBUG
add_allocated_tensor(alloc, tensor);
size_t cur_max = (char*)addr - (char*)alloc->data + size;
size_t cur_max = (char*)addr - (char*)alloc->base + size;
if (cur_max > alloc->max_size) {
printf("max_size = %.2f MB: tensors: ", cur_max / 1024.0 / 1024.0);
for (int i = 0; i < 1024; i++) {
@ -168,10 +168,6 @@ static void ggml_tallocr_free_tensor(ggml_tallocr_t alloc, struct ggml_tensor *
size = aligned_offset(NULL, size, alloc->alignment);
AT_PRINTF("%s: freeing %s at %p (%zu bytes) - n_free_blocks = %d\n", __func__, tensor->name, ptr, size, alloc->n_free_blocks);
if (!alloc->measure) {
ggml_backend_buffer_free_tensor(alloc->buffer, tensor);
}
#ifdef GGML_ALLOCATOR_DEBUG
remove_allocated_tensor(alloc, tensor);
#endif
@ -237,7 +233,7 @@ void ggml_tallocr_reset(ggml_tallocr_t alloc) {
}
ggml_tallocr_t ggml_tallocr_new(void * data, size_t size, size_t alignment) {
struct ggml_backend_buffer * buffer = ggml_backend_cpu_buffer_from_ptr(NULL, data, size);
struct ggml_backend_buffer * buffer = ggml_backend_cpu_buffer_from_ptr(data, size);
ggml_tallocr_t alloc = (ggml_tallocr_t)malloc(sizeof(struct ggml_tallocr));
@ -449,17 +445,15 @@ static ggml_tallocr_t node_tallocr(ggml_gallocr_t galloc, struct ggml_tensor * n
static void init_view(ggml_gallocr_t galloc, struct ggml_tensor * view, bool update_backend) {
ggml_tallocr_t alloc = node_tallocr(galloc, view);
//printf("init_view: %s from src %s\n", view->name, view->view_src->name);
GGML_ASSERT(view->view_src != NULL && view->view_src->data != NULL);
if (update_backend) {
view->backend = view->view_src->backend;
}
view->buffer = view->view_src->buffer;
// views are initialized in the alloc buffer rather than the view_src buffer
view->buffer = alloc->buffer;
view->data = (char *)view->view_src->data + view->view_offs;
// FIXME: the view should be initialized by the owning buffer, but currently this breaks the CUDA backend
// due to the ggml_tensor_extra_gpu ring buffer overwriting the KV cache extras
assert(ggml_tallocr_is_measure(alloc) || !view->buffer || view->buffer->backend == alloc->buffer->backend);
assert(ggml_tallocr_is_measure(alloc) || !view->buffer || view->buffer->buft == alloc->buffer->buft);
if (!alloc->measure) {
ggml_backend_buffer_init_tensor(alloc->buffer, view);
@ -741,6 +735,10 @@ void ggml_allocr_set_parse_seq(ggml_allocr_t alloc, const int * list, int n) {
}
void ggml_allocr_free(ggml_allocr_t alloc) {
if (alloc == NULL) {
return;
}
ggml_gallocr_free(alloc->galloc);
ggml_tallocr_free(alloc->talloc);
free(alloc);
@ -765,3 +763,48 @@ size_t ggml_allocr_max_size(ggml_allocr_t alloc) {
size_t ggml_allocr_alloc_graph(ggml_allocr_t alloc, struct ggml_cgraph * graph) {
return ggml_gallocr_alloc_graph(alloc->galloc, alloc->talloc, graph);
}
// utils
ggml_backend_buffer_t ggml_backend_alloc_ctx_tensors_from_buft(struct ggml_context * ctx, ggml_backend_buffer_type_t buft) {
GGML_ASSERT(ggml_get_no_alloc(ctx) == true);
size_t alignment = ggml_backend_buft_get_alignment(buft);
size_t nbytes = 0;
for (struct ggml_tensor * t = ggml_get_first_tensor(ctx); t != NULL; t = ggml_get_next_tensor(ctx, t)) {
if (t->data == NULL && t->view_src == NULL) {
nbytes += GGML_PAD(ggml_backend_buft_get_alloc_size(buft, t), alignment);
}
}
if (nbytes == 0) {
// all the tensors in the context are already allocated
return NULL;
}
ggml_backend_buffer_t buffer = ggml_backend_buft_alloc_buffer(buft, nbytes);
ggml_tallocr_t tallocr = ggml_tallocr_new_from_buffer(buffer);
for (struct ggml_tensor * t = ggml_get_first_tensor(ctx); t != NULL; t = ggml_get_next_tensor(ctx, t)) {
if (t->data == NULL) {
if (t->view_src == NULL) {
ggml_tallocr_alloc(tallocr, t);
} else {
ggml_backend_view_init(buffer, t);
}
} else {
if (t->view_src != NULL) {
// view of a pre-allocated tensor
ggml_backend_view_init(buffer, t);
}
}
}
ggml_tallocr_free(tallocr);
return buffer;
}
ggml_backend_buffer_t ggml_backend_alloc_ctx_tensors(struct ggml_context * ctx, ggml_backend_t backend) {
return ggml_backend_alloc_ctx_tensors_from_buft(ctx, ggml_backend_get_default_buffer_type(backend));
}

9
ggml-alloc.h
View file

@ -8,6 +8,7 @@ extern "C" {
struct ggml_backend;
struct ggml_backend_buffer;
struct ggml_backend_buffer_type;
//
// Legacy API
@ -42,7 +43,7 @@ GGML_API size_t ggml_allocr_alloc_graph(ggml_allocr_t alloc, struct ggml_cgraph
// ggml-backend v2 API
//
// Seperate tensor and graph allocator objects
// Separate tensor and graph allocator objects
// This is necessary for multi-backend allocation because the graph allocator needs to use multiple tensor allocators
// The original API is kept as a wrapper around the new API
@ -80,6 +81,12 @@ GGML_API void ggml_gallocr_alloc_graph_n(
struct ggml_hash_set hash_set,
ggml_tallocr_t * hash_node_talloc);
// Utils
// Create a buffer and allocate all the tensors in a ggml_context
GGML_API struct ggml_backend_buffer * ggml_backend_alloc_ctx_tensors_from_buft(struct ggml_context * ctx, struct ggml_backend_buffer_type * buft);
GGML_API struct ggml_backend_buffer * ggml_backend_alloc_ctx_tensors(struct ggml_context * ctx, struct ggml_backend * backend);
#ifdef __cplusplus
}
#endif

71
ggml-backend-impl.h
View file

@ -12,31 +12,54 @@ extern "C" {
// Backend buffer
//
// buffer type
typedef void * ggml_backend_buffer_type_context_t;
struct ggml_backend_buffer_type_i {
ggml_backend_buffer_t (*alloc_buffer) (ggml_backend_buffer_type_t buft, size_t size);
size_t (*get_alignment) (ggml_backend_buffer_type_t buft); // tensor alignment
size_t (*get_alloc_size) (ggml_backend_buffer_type_t buft, struct ggml_tensor * tensor); // data size needed to allocate the tensor, including padding
bool (*supports_backend)(ggml_backend_buffer_type_t buft, ggml_backend_t backend); // check if the buffer type is usable by the backend
// check if tensor data is in host memory
// should be equivalent to supports_backend(buft, ggml_backend_cpu_init())
bool (*is_host) (ggml_backend_buffer_type_t buft);
};
struct ggml_backend_buffer_type {
struct ggml_backend_buffer_type_i iface;
ggml_backend_buffer_type_context_t context;
};
// buffer
typedef void * ggml_backend_buffer_context_t;
struct ggml_backend_buffer_i {
void (*free_buffer) (ggml_backend_buffer_t buffer);
void * (*get_base) (ggml_backend_buffer_t buffer); // get base pointer
size_t (*get_alloc_size)(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor); // pre-allocation callback
void (*init_tensor) (ggml_backend_buffer_t buffer, struct ggml_tensor * tensor); // post-allocation callback
void (*free_tensor) (ggml_backend_buffer_t buffer, struct ggml_tensor * tensor); // pre-free callback
void (*free_buffer) (ggml_backend_buffer_t buffer);
//void (*reset) (ggml_backend_buffer_t buffer); // reset any internal state due to tensor initialization, such as tensor extras
void * (*get_base) (ggml_backend_buffer_t buffer);
void (*init_tensor) (ggml_backend_buffer_t buffer, struct ggml_tensor * tensor);
void (*set_tensor) (ggml_backend_buffer_t buffer, struct ggml_tensor * tensor, const void * data, size_t offset, size_t size);
void (*get_tensor) (ggml_backend_buffer_t buffer, const struct ggml_tensor * tensor, void * data, size_t offset, size_t size);
// (optional) copy tensor between different buffer-type, allow for single-copy tranfers
void (*cpy_tensor_from)(ggml_backend_buffer_t buffer, struct ggml_tensor * src, struct ggml_tensor * dst);
void (*cpy_tensor_to) (ggml_backend_buffer_t buffer, struct ggml_tensor * src, struct ggml_tensor * dst);
void (*clear) (ggml_backend_buffer_t buffer, uint8_t value);
};
struct ggml_backend_buffer {
struct ggml_backend_buffer_i iface;
ggml_backend_t backend;
struct ggml_backend_buffer_i iface;
ggml_backend_buffer_type_t buft;
ggml_backend_buffer_context_t context;
size_t size;
};
GGML_API ggml_backend_buffer_t ggml_backend_buffer_init(
struct ggml_backend * backend,
ggml_backend_buffer_t ggml_backend_buffer_init(
ggml_backend_buffer_type_t buft,
struct ggml_backend_buffer_i iface,
ggml_backend_buffer_context_t context,
size_t size);
//
// Backend
//
@ -49,20 +72,17 @@ extern "C" {
void (*free)(ggml_backend_t backend);
// buffer allocation
ggml_backend_buffer_t (*alloc_buffer)(ggml_backend_t backend, size_t size);
ggml_backend_buffer_type_t (*get_default_buffer_type)(ggml_backend_t backend);
// get buffer alignment
size_t (*get_alignment)(ggml_backend_t backend);
// tensor data access
// these functions can be asynchronous, helper functions are provided for synchronous access that automatically call synchronize
// (optional) asynchroneous tensor data access
void (*set_tensor_async)(ggml_backend_t backend, struct ggml_tensor * tensor, const void * data, size_t offset, size_t size);
void (*get_tensor_async)(ggml_backend_t backend, const struct ggml_tensor * tensor, void * data, size_t offset, size_t size);
void (*synchronize) (ggml_backend_t backend);
// (optional) copy tensor between different backends, allow for single-copy tranfers
void (*cpy_tensor_from)(ggml_backend_t backend, struct ggml_tensor * src, struct ggml_tensor * dst);
void (*cpy_tensor_to) (ggml_backend_t backend, struct ggml_tensor * src, struct ggml_tensor * dst);
// (optional) asynchroneous tensor copy
void (*cpy_tensor_from_async)(ggml_backend_t backend, struct ggml_tensor * src, struct ggml_tensor * dst);
void (*cpy_tensor_to_async) (ggml_backend_t backend, struct ggml_tensor * src, struct ggml_tensor * dst);
void (*synchronize)(ggml_backend_t backend);
// compute graph with a plan
ggml_backend_graph_plan_t (*graph_plan_create) (ggml_backend_t backend, struct ggml_cgraph * cgraph);
@ -82,6 +102,15 @@ extern "C" {
ggml_backend_context_t context;
};
//
// Backend registry
//
typedef ggml_backend_t (*ggml_backend_init_fn)(const char * params, void * user_data);
void ggml_backend_register(const char * name, ggml_backend_init_fn init_fn, ggml_backend_buffer_type_t default_buffer_type, void * user_data);
#ifdef __cplusplus
}
#endif

839
ggml-backend.c
View file

File diff suppressed because it is too large Load diff

86
ggml-backend.h
View file

@ -7,41 +7,47 @@
extern "C" {
#endif
typedef struct ggml_backend_buffer_type * ggml_backend_buffer_type_t;
typedef struct ggml_backend_buffer * ggml_backend_buffer_t;
typedef struct ggml_backend * ggml_backend_t;
typedef void * ggml_backend_graph_plan_t;
//
// Backend buffer
//
struct ggml_backend_buffer;
typedef struct ggml_backend_buffer * ggml_backend_buffer_t;
// buffer type
GGML_API ggml_backend_buffer_t ggml_backend_buft_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size);
GGML_API size_t ggml_backend_buft_get_alignment (ggml_backend_buffer_type_t buft);
GGML_API size_t ggml_backend_buft_get_alloc_size(ggml_backend_buffer_type_t buft, struct ggml_tensor * tensor);
GGML_API bool ggml_backend_buft_supports_backend(ggml_backend_buffer_type_t buft, ggml_backend_t backend);
GGML_API bool ggml_backend_buft_is_host (ggml_backend_buffer_type_t buft);
// backend buffer functions
// buffer
GGML_API void ggml_backend_buffer_free (ggml_backend_buffer_t buffer);
GGML_API size_t ggml_backend_buffer_get_alignment (ggml_backend_buffer_t buffer);
GGML_API void * ggml_backend_buffer_get_base (ggml_backend_buffer_t buffer);
GGML_API size_t ggml_backend_buffer_get_size (ggml_backend_buffer_t buffer);
GGML_API size_t ggml_backend_buffer_get_alloc_size(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor);
GGML_API void ggml_backend_buffer_init_tensor (ggml_backend_buffer_t buffer, struct ggml_tensor * tensor);
GGML_API void ggml_backend_buffer_free_tensor (ggml_backend_buffer_t buffer, struct ggml_tensor * tensor);
GGML_API size_t ggml_backend_buffer_get_alignment (ggml_backend_buffer_t buffer);
GGML_API size_t ggml_backend_buffer_get_alloc_size(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor);
GGML_API void ggml_backend_buffer_clear (ggml_backend_buffer_t buffer, uint8_t value);
GGML_API bool ggml_backend_buffer_is_host (ggml_backend_buffer_t buffer);
GGML_API ggml_backend_buffer_type_t ggml_backend_buffer_type(ggml_backend_buffer_t buffer);
//
// Backend
//
struct ggml_backend;
typedef struct ggml_backend * ggml_backend_t;
typedef void * ggml_backend_graph_plan_t;
GGML_API ggml_backend_t ggml_get_backend(const struct ggml_tensor * tensor);
GGML_API const char * ggml_backend_name(ggml_backend_t backend);
GGML_API void ggml_backend_free(ggml_backend_t backend);
GGML_API ggml_backend_buffer_t ggml_backend_alloc_buffer(ggml_backend_t backend, size_t size);
GGML_API ggml_backend_buffer_type_t ggml_backend_get_default_buffer_type(ggml_backend_t backend);
GGML_API ggml_backend_buffer_t ggml_backend_alloc_buffer(ggml_backend_t backend, size_t size);
GGML_API size_t ggml_backend_get_alignment(ggml_backend_t backend);
GGML_API size_t ggml_backend_get_alignment(ggml_backend_t backend);
GGML_API void ggml_backend_tensor_set_async( struct ggml_tensor * tensor, const void * data, size_t offset, size_t size);
GGML_API void ggml_backend_tensor_get_async(const struct ggml_tensor * tensor, void * data, size_t offset, size_t size);
GGML_API void ggml_backend_tensor_set_async(ggml_backend_t backend, struct ggml_tensor * tensor, const void * data, size_t offset, size_t size);
GGML_API void ggml_backend_tensor_get_async(ggml_backend_t backend, const struct ggml_tensor * tensor, void * data, size_t offset, size_t size);
GGML_API void ggml_backend_tensor_set( struct ggml_tensor * tensor, const void * data, size_t offset, size_t size);
GGML_API void ggml_backend_tensor_get(const struct ggml_tensor * tensor, void * data, size_t offset, size_t size);
@ -57,6 +63,7 @@ extern "C" {
// tensor copy between different backends
GGML_API void ggml_backend_tensor_copy(struct ggml_tensor * src, struct ggml_tensor * dst);
GGML_API void ggml_backend_tensor_copy_async(ggml_backend_t backend, struct ggml_tensor * src, struct ggml_tensor * dst); // automatic fallback to sync copy
//
// CPU backend
@ -68,8 +75,27 @@ extern "C" {
GGML_API void ggml_backend_cpu_set_n_threads(ggml_backend_t backend_cpu, int n_threads);
// Create a backend buffer from an existing pointer
GGML_API ggml_backend_buffer_t ggml_backend_cpu_buffer_from_ptr(ggml_backend_t backend_cpu, void * ptr, size_t size);
GGML_API ggml_backend_buffer_t ggml_backend_cpu_buffer_from_ptr(void * ptr, size_t size);
GGML_API ggml_backend_buffer_type_t ggml_backend_cpu_buffer_type(void);
#ifdef GGML_USE_CPU_HBM
GGML_API ggml_backend_buffer_type_t ggml_backend_cpu_hbm_buffer_type(void);
#endif
//
// Backend registry
//
// The backend registry is a registry of all the available backends, and allows initializing backends in a generic way
GGML_API size_t ggml_backend_reg_get_count(void);
GGML_API size_t ggml_backend_reg_find_by_name(const char * name);
GGML_API ggml_backend_t ggml_backend_reg_init_backend_from_str(const char * backend_str); // str is name[:params]
GGML_API const char * ggml_backend_reg_get_name(size_t i);
GGML_API ggml_backend_t ggml_backend_reg_init_backend(size_t i, const char * params); // params is backend-specific
GGML_API ggml_backend_buffer_type_t ggml_backend_reg_get_default_buffer_type(size_t i);
GGML_API ggml_backend_buffer_t ggml_backend_reg_alloc_buffer(size_t i, size_t size);
//
// Backend scheduler
@ -131,6 +157,32 @@ extern "C" {
ggml_backend_sched_t sched,
struct ggml_cgraph * graph);
//
// Utils
//
struct ggml_backend_graph_copy {
ggml_backend_buffer_t buffer;
struct ggml_context * ctx_allocated;
struct ggml_context * ctx_unallocated;
struct ggml_cgraph * graph;
};
// Copy a graph to a different backend
GGML_API struct ggml_backend_graph_copy ggml_backend_graph_copy(ggml_backend_t backend, struct ggml_cgraph * graph);
GGML_API void ggml_backend_graph_copy_free(struct ggml_backend_graph_copy copy);
typedef bool (*ggml_backend_eval_callback)(int node_index, struct ggml_tensor * t1, struct ggml_tensor * t2, void * user_data);
// Compare the output of two backends
GGML_API void ggml_backend_compare_graph_backend(ggml_backend_t backend1, ggml_backend_t backend2, struct ggml_cgraph * graph, ggml_backend_eval_callback callback, void * user_data);
// Tensor initialization
GGML_API void ggml_backend_tensor_alloc(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor, void * addr);
GGML_API void ggml_backend_view_init(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor);
#ifdef __cplusplus
}
#endif

3407
ggml-cuda.cu
View file

File diff suppressed because it is too large Load diff

10
ggml-cuda.h
View file

@ -49,7 +49,15 @@ GGML_API int ggml_cuda_get_device_count(void);
GGML_API void ggml_cuda_get_device_description(int device, char * description, size_t description_size);
// backend API
GGML_API ggml_backend_t ggml_backend_cuda_init(void); // TODO: take a list of devices to use
GGML_API ggml_backend_t ggml_backend_cuda_init(int device);
GGML_API bool ggml_backend_is_cuda(ggml_backend_t backend);
GGML_API int ggml_backend_cuda_get_device(ggml_backend_t backend);
GGML_API ggml_backend_buffer_type_t ggml_backend_cuda_buffer_type(int device);
// pinned host buffer for use with CPU backend for faster copies between CPU and GPU
GGML_API ggml_backend_buffer_type_t ggml_backend_cuda_host_buffer_type(void);
#ifdef __cplusplus
}

2
ggml-impl.h
View file

@ -232,7 +232,7 @@ bool ggml_hash_contains (const struct ggml_hash_set hash_set, struct ggml
// returns GGML_HASHTABLE_FULL if table is full, otherwise the current index of the key or where it should be inserted
size_t ggml_hash_find (const struct ggml_hash_set hash_set, struct ggml_tensor * key);
// returns GGML_HAHSHTABLE_ALREADY_EXISTS if key already exists, index otherwise, asserts if table is full
// returns GGML_HASHTABLE_ALREADY_EXISTS if key already exists, index otherwise, asserts if table is full
size_t ggml_hash_insert ( struct ggml_hash_set hash_set, struct ggml_tensor * key);
// return index, asserts if table is full

9
ggml-metal.h
View file

@ -98,8 +98,17 @@ GGML_API ggml_backend_t ggml_backend_metal_init(void);
GGML_API bool ggml_backend_is_metal(ggml_backend_t backend);
GGML_API ggml_backend_buffer_t ggml_backend_metal_buffer_from_ptr(void * data, size_t size, size_t max_size);
GGML_API void ggml_backend_metal_set_n_cb(ggml_backend_t backend, int n_cb);
GGML_API ggml_backend_buffer_type_t ggml_backend_metal_buffer_type(void);
// helper to check if the device supports a specific family
// ideally, the user code should be doing these checks
// ref: https://developer.apple.com/metal/Metal-Feature-Set-Tables.pdf
GGML_API bool ggml_backend_metal_supports_family(ggml_backend_t backend, int family);
#ifdef __cplusplus
}
#endif

1417
ggml-metal.m
View file

File diff suppressed because it is too large Load diff

2668
ggml-metal.metal
View file

File diff suppressed because it is too large Load diff

12
ggml-opencl.cpp
View file

@ -1,20 +1,18 @@
#include "ggml.h"
#include "ggml-opencl.h"
#include <array>
#include <atomic>
#include <cstdio>
#include <cstdlib>
#include <cstring>
#include <limits>
#include <sstream>
#include <vector>
#include <limits>
#define CL_TARGET_OPENCL_VERSION 110
#include <clblast.h>
#include <stdlib.h>
#include <stdio.h>
#include <string.h>
#include "ggml.h"
#if defined(_MSC_VER)
#pragma warning(disable: 4244 4267) // possible loss of data
#endif

371
ggml-quants.c
View file

@ -407,6 +407,18 @@ inline static ggml_int8x16x4_t ggml_vld1q_s8_x4(const int8_t * ptr) {
#define ggml_vld1q_s8_x4 vld1q_s8_x4
#endif
#if !defined(__ARM_FEATURE_DOTPROD)
inline static int32x4_t vdotq_s32(int32x4_t acc, int8x16_t a, int8x16_t b) {
const int16x8_t p0 = vmull_s8(vget_low_s8 (a), vget_low_s8 (b));
const int16x8_t p1 = vmull_s8(vget_high_s8(a), vget_high_s8(b));
return vaddq_s32(acc, vaddq_s32(vpaddlq_s16(p0), vpaddlq_s16(p1)));
}
#endif
#endif
#if defined(__ARM_NEON) || defined(__wasm_simd128__)
@ -2468,32 +2480,12 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * restrict s, const void * restrict vx,
const int8x16_t v1_1l = vld1q_s8(y1->qs);
const int8x16_t v1_1h = vld1q_s8(y1->qs + 16);
#if defined(__ARM_FEATURE_DOTPROD)
// dot product into int32x4_t
const int32x4_t p_0 = vdotq_s32(vdotq_s32(vdupq_n_s32(0), v0_0ls, v1_0l), v0_0hs, v1_0h);
const int32x4_t p_1 = vdotq_s32(vdotq_s32(vdupq_n_s32(0), v0_1ls, v1_1l), v0_1hs, v1_1h);
sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(p_0), GGML_FP16_TO_FP32(x0->d)*GGML_FP16_TO_FP32(y0->d));
sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(p_1), GGML_FP16_TO_FP32(x1->d)*GGML_FP16_TO_FP32(y1->d));
#else
const int16x8_t pl0l = vmull_s8(vget_low_s8 (v0_0ls), vget_low_s8 (v1_0l));
const int16x8_t pl0h = vmull_s8(vget_high_s8(v0_0ls), vget_high_s8(v1_0l));
const int16x8_t ph0l = vmull_s8(vget_low_s8 (v0_0hs), vget_low_s8 (v1_0h));
const int16x8_t ph0h = vmull_s8(vget_high_s8(v0_0hs), vget_high_s8(v1_0h));
const int16x8_t pl1l = vmull_s8(vget_low_s8 (v0_1ls), vget_low_s8 (v1_1l));
const int16x8_t pl1h = vmull_s8(vget_high_s8(v0_1ls), vget_high_s8(v1_1l));
const int16x8_t ph1l = vmull_s8(vget_low_s8 (v0_1hs), vget_low_s8 (v1_1h));
const int16x8_t ph1h = vmull_s8(vget_high_s8(v0_1hs), vget_high_s8(v1_1h));
const int32x4_t pl0 = vaddq_s32(vpaddlq_s16(pl0l), vpaddlq_s16(pl0h));
const int32x4_t ph0 = vaddq_s32(vpaddlq_s16(ph0l), vpaddlq_s16(ph0h));
const int32x4_t pl1 = vaddq_s32(vpaddlq_s16(pl1l), vpaddlq_s16(pl1h));
const int32x4_t ph1 = vaddq_s32(vpaddlq_s16(ph1l), vpaddlq_s16(ph1h));
sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(vaddq_s32(pl0, ph0)), GGML_FP16_TO_FP32(x0->d)*GGML_FP16_TO_FP32(y0->d));
sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(vaddq_s32(pl1, ph1)), GGML_FP16_TO_FP32(x1->d)*GGML_FP16_TO_FP32(y1->d));
#endif
}
*s = vaddvq_f32(sumv0) + vaddvq_f32(sumv1);
@ -2776,32 +2768,12 @@ void ggml_vec_dot_q4_1_q8_1(const int n, float * restrict s, const void * restri
const int8x16_t v1_1l = vld1q_s8(y1->qs);
const int8x16_t v1_1h = vld1q_s8(y1->qs + 16);
#if defined(__ARM_FEATURE_DOTPROD)
// dot product into int32x4_t
const int32x4_t p_0 = vdotq_s32(vdotq_s32(vdupq_n_s32(0), v0_0l, v1_0l), v0_0h, v1_0h);
const int32x4_t p_1 = vdotq_s32(vdotq_s32(vdupq_n_s32(0), v0_1l, v1_1l), v0_1h, v1_1h);
sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(p_0), GGML_FP16_TO_FP32(x0->d)*y0->d);
sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(p_1), GGML_FP16_TO_FP32(x1->d)*y1->d);
#else
const int16x8_t pl0l = vmull_s8(vget_low_s8 (v0_0l), vget_low_s8 (v1_0l));
const int16x8_t pl0h = vmull_s8(vget_high_s8(v0_0l), vget_high_s8(v1_0l));
const int16x8_t ph0l = vmull_s8(vget_low_s8 (v0_0h), vget_low_s8 (v1_0h));
const int16x8_t ph0h = vmull_s8(vget_high_s8(v0_0h), vget_high_s8(v1_0h));
const int16x8_t pl1l = vmull_s8(vget_low_s8 (v0_1l), vget_low_s8 (v1_1l));
const int16x8_t pl1h = vmull_s8(vget_high_s8(v0_1l), vget_high_s8(v1_1l));
const int16x8_t ph1l = vmull_s8(vget_low_s8 (v0_1h), vget_low_s8 (v1_1h));
const int16x8_t ph1h = vmull_s8(vget_high_s8(v0_1h), vget_high_s8(v1_1h));
const int32x4_t pl0 = vaddq_s32(vpaddlq_s16(pl0l), vpaddlq_s16(pl0h));
const int32x4_t ph0 = vaddq_s32(vpaddlq_s16(ph0l), vpaddlq_s16(ph0h));
const int32x4_t pl1 = vaddq_s32(vpaddlq_s16(pl1l), vpaddlq_s16(pl1h));
const int32x4_t ph1 = vaddq_s32(vpaddlq_s16(ph1l), vpaddlq_s16(ph1h));
sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(vaddq_s32(pl0, ph0)), GGML_FP16_TO_FP32(x0->d)*y0->d);
sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(vaddq_s32(pl1, ph1)), GGML_FP16_TO_FP32(x1->d)*y1->d);
#endif
}
*s = vaddvq_f32(sumv0) + vaddvq_f32(sumv1) + summs;
@ -2963,32 +2935,12 @@ void ggml_vec_dot_q5_0_q8_0(const int n, float * restrict s, const void * restri
const int8x16_t v1_1l = vld1q_s8(y1->qs);
const int8x16_t v1_1h = vld1q_s8(y1->qs + 16);
#if defined(__ARM_FEATURE_DOTPROD)
sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(vaddq_s32(
vdotq_s32(vdupq_n_s32(0), v0_0lf, v1_0l),
vdotq_s32(vdupq_n_s32(0), v0_0hf, v1_0h))), GGML_FP16_TO_FP32(x0->d)*GGML_FP16_TO_FP32(y0->d));
sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(vaddq_s32(
vdotq_s32(vdupq_n_s32(0), v0_1lf, v1_1l),
vdotq_s32(vdupq_n_s32(0), v0_1hf, v1_1h))), GGML_FP16_TO_FP32(x1->d)*GGML_FP16_TO_FP32(y1->d));
#else
const int16x8_t pl0l = vmull_s8(vget_low_s8 (v0_0lf), vget_low_s8 (v1_0l));
const int16x8_t pl0h = vmull_s8(vget_high_s8(v0_0lf), vget_high_s8(v1_0l));
const int16x8_t ph0l = vmull_s8(vget_low_s8 (v0_0hf), vget_low_s8 (v1_0h));
const int16x8_t ph0h = vmull_s8(vget_high_s8(v0_0hf), vget_high_s8(v1_0h));
const int16x8_t pl1l = vmull_s8(vget_low_s8 (v0_1lf), vget_low_s8 (v1_1l));
const int16x8_t pl1h = vmull_s8(vget_high_s8(v0_1lf), vget_high_s8(v1_1l));
const int16x8_t ph1l = vmull_s8(vget_low_s8 (v0_1hf), vget_low_s8 (v1_1h));
const int16x8_t ph1h = vmull_s8(vget_high_s8(v0_1hf), vget_high_s8(v1_1h));
const int32x4_t pl0 = vaddq_s32(vpaddlq_s16(pl0l), vpaddlq_s16(pl0h));
const int32x4_t ph0 = vaddq_s32(vpaddlq_s16(ph0l), vpaddlq_s16(ph0h));
const int32x4_t pl1 = vaddq_s32(vpaddlq_s16(pl1l), vpaddlq_s16(pl1h));
const int32x4_t ph1 = vaddq_s32(vpaddlq_s16(ph1l), vpaddlq_s16(ph1h));
sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(vaddq_s32(pl0, ph0)), GGML_FP16_TO_FP32(x0->d)*GGML_FP16_TO_FP32(y0->d));
sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(vaddq_s32(pl1, ph1)), GGML_FP16_TO_FP32(x1->d)*GGML_FP16_TO_FP32(y1->d));
#endif
}
*s = vaddvq_f32(sumv0) + vaddvq_f32(sumv1);
@ -3114,7 +3066,7 @@ void ggml_vec_dot_q5_0_q8_0(const int n, float * restrict s, const void * restri
size_t vl = __riscv_vsetvl_e8m1(qk/2);
// These tempory registers are for masking and shift operations
// These temporary registers are for masking and shift operations
vuint32m2_t vt_1 = __riscv_vid_v_u32m2(vl);
vuint32m2_t vt_2 = __riscv_vsll_vv_u32m2(__riscv_vmv_v_x_u32m2(1, vl), vt_1, vl);
@ -3275,32 +3227,12 @@ void ggml_vec_dot_q5_1_q8_1(const int n, float * restrict s, const void * restri
const int8x16_t v1_1l = vld1q_s8(y1->qs);
const int8x16_t v1_1h = vld1q_s8(y1->qs + 16);
#if defined(__ARM_FEATURE_DOTPROD)
sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(vaddq_s32(
vdotq_s32(vdupq_n_s32(0), v0_0lf, v1_0l),
vdotq_s32(vdupq_n_s32(0), v0_0hf, v1_0h))), GGML_FP16_TO_FP32(x0->d)*y0->d);
sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(vaddq_s32(
vdotq_s32(vdupq_n_s32(0), v0_1lf, v1_1l),
vdotq_s32(vdupq_n_s32(0), v0_1hf, v1_1h))), GGML_FP16_TO_FP32(x1->d)*y1->d);
#else
const int16x8_t pl0l = vmull_s8(vget_low_s8 (v0_0lf), vget_low_s8 (v1_0l));
const int16x8_t pl0h = vmull_s8(vget_high_s8(v0_0lf), vget_high_s8(v1_0l));
const int16x8_t ph0l = vmull_s8(vget_low_s8 (v0_0hf), vget_low_s8 (v1_0h));
const int16x8_t ph0h = vmull_s8(vget_high_s8(v0_0hf), vget_high_s8(v1_0h));
const int16x8_t pl1l = vmull_s8(vget_low_s8 (v0_1lf), vget_low_s8 (v1_1l));
const int16x8_t pl1h = vmull_s8(vget_high_s8(v0_1lf), vget_high_s8(v1_1l));
const int16x8_t ph1l = vmull_s8(vget_low_s8 (v0_1hf), vget_low_s8 (v1_1h));
const int16x8_t ph1h = vmull_s8(vget_high_s8(v0_1hf), vget_high_s8(v1_1h));
const int32x4_t pl0 = vaddq_s32(vpaddlq_s16(pl0l), vpaddlq_s16(pl0h));
const int32x4_t ph0 = vaddq_s32(vpaddlq_s16(ph0l), vpaddlq_s16(ph0h));
const int32x4_t pl1 = vaddq_s32(vpaddlq_s16(pl1l), vpaddlq_s16(pl1h));
const int32x4_t ph1 = vaddq_s32(vpaddlq_s16(ph1l), vpaddlq_s16(ph1h));
sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(vaddq_s32(pl0, ph0)), GGML_FP16_TO_FP32(x0->d)*y0->d);
sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(vaddq_s32(pl1, ph1)), GGML_FP16_TO_FP32(x1->d)*y1->d);
#endif
}
*s = vaddvq_f32(sumv0) + vaddvq_f32(sumv1) + summs0 + summs1;
@ -3550,7 +3482,6 @@ void ggml_vec_dot_q8_0_q8_0(const int n, float * restrict s, const void * restri
const int8x16_t y1_0 = vld1q_s8(y1->qs);
const int8x16_t y1_1 = vld1q_s8(y1->qs + 16);
#if defined(__ARM_FEATURE_DOTPROD)
sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(vaddq_s32(
vdotq_s32(vdupq_n_s32(0), x0_0, y0_0),
vdotq_s32(vdupq_n_s32(0), x0_1, y0_1))), GGML_FP16_TO_FP32(x0->d)*GGML_FP16_TO_FP32(y0->d));
@ -3558,26 +3489,6 @@ void ggml_vec_dot_q8_0_q8_0(const int n, float * restrict s, const void * restri
sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(vaddq_s32(
vdotq_s32(vdupq_n_s32(0), x1_0, y1_0),
vdotq_s32(vdupq_n_s32(0), x1_1, y1_1))), GGML_FP16_TO_FP32(x1->d)*GGML_FP16_TO_FP32(y1->d));
#else
const int16x8_t p0_0 = vmull_s8(vget_low_s8 (x0_0), vget_low_s8 (y0_0));
const int16x8_t p0_1 = vmull_s8(vget_high_s8(x0_0), vget_high_s8(y0_0));
const int16x8_t p0_2 = vmull_s8(vget_low_s8 (x0_1), vget_low_s8 (y0_1));
const int16x8_t p0_3 = vmull_s8(vget_high_s8(x0_1), vget_high_s8(y0_1));
const int16x8_t p1_0 = vmull_s8(vget_low_s8 (x1_0), vget_low_s8 (y1_0));
const int16x8_t p1_1 = vmull_s8(vget_high_s8(x1_0), vget_high_s8(y1_0));
const int16x8_t p1_2 = vmull_s8(vget_low_s8 (x1_1), vget_low_s8 (y1_1));
const int16x8_t p1_3 = vmull_s8(vget_high_s8(x1_1), vget_high_s8(y1_1));
const int32x4_t p0 = vaddq_s32(vpaddlq_s16(p0_0), vpaddlq_s16(p0_1));
const int32x4_t p1 = vaddq_s32(vpaddlq_s16(p0_2), vpaddlq_s16(p0_3));
const int32x4_t p2 = vaddq_s32(vpaddlq_s16(p1_0), vpaddlq_s16(p1_1));
const int32x4_t p3 = vaddq_s32(vpaddlq_s16(p1_2), vpaddlq_s16(p1_3));
sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(vaddq_s32(p0, p1)), GGML_FP16_TO_FP32(x0->d)*GGML_FP16_TO_FP32(y0->d));
sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(vaddq_s32(p2, p3)), GGML_FP16_TO_FP32(x1->d)*GGML_FP16_TO_FP32(y1->d));
#endif
}
*s = vaddvq_f32(sumv0) + vaddvq_f32(sumv1);
@ -3650,12 +3561,10 @@ void ggml_vec_dot_q2_K_q8_K(const int n, float * restrict s, const void * restri
const int nb = n / QK_K;
#ifdef __ARM_NEON
const uint8x16_t m3 = vdupq_n_u8(0x3);
const uint8x16_t m4 = vdupq_n_u8(0xF);
#if defined(__ARM_FEATURE_DOTPROD)
const int32x4_t vzero = vdupq_n_s32(0);
#endif
const int32x4_t vzero = vdupq_n_s32(0);
ggml_int8x16x2_t q2bytes;
uint8_t aux[16];
@ -3663,7 +3572,6 @@ void ggml_vec_dot_q2_K_q8_K(const int n, float * restrict s, const void * restri
float sum = 0;
for (int i = 0; i < nb; ++i) {
const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
const float dmin = -y[i].d * GGML_FP16_TO_FP32(x[i].dmin);
@ -3677,7 +3585,7 @@ void ggml_vec_dot_q2_K_q8_K(const int n, float * restrict s, const void * restri
const uint8x16_t mins = vshrq_n_u8(mins_and_scales, 4);
const ggml_int16x8x2_t q8sums = ggml_vld1q_s16_x2(y[i].bsums);
const ggml_int16x8x2_t mins16 = {vreinterpretq_s16_u16(vmovl_u8(vget_low_u8(mins))), vreinterpretq_s16_u16(vmovl_u8(vget_high_u8(mins)))};
const ggml_int16x8x2_t mins16 = {{vreinterpretq_s16_u16(vmovl_u8(vget_low_u8(mins))), vreinterpretq_s16_u16(vmovl_u8(vget_high_u8(mins)))}};
const int32x4_t s0 = vaddq_s32(vmull_s16(vget_low_s16 (mins16.val[0]), vget_low_s16 (q8sums.val[0])),
vmull_s16(vget_high_s16(mins16.val[0]), vget_high_s16(q8sums.val[0])));
const int32x4_t s1 = vaddq_s32(vmull_s16(vget_low_s16 (mins16.val[1]), vget_low_s16 (q8sums.val[1])),
@ -3689,20 +3597,9 @@ void ggml_vec_dot_q2_K_q8_K(const int n, float * restrict s, const void * restri
// We use this macro instead of a function call because for some reason
// the code runs 2-3% slower, even if the function is declared inline
#if defined(__ARM_FEATURE_DOTPROD)
#define MULTIPLY_ACCUM_WITH_SCALE(index)\
isum += vaddvq_s32(vdotq_s32(vzero, q2bytes.val[0], q8bytes.val[0])) * aux[is+(index)];\
isum += vaddvq_s32(vdotq_s32(vzero, q2bytes.val[1], q8bytes.val[1])) * aux[is+1+(index)];
#else
#define MULTIPLY_ACCUM_WITH_SCALE(index)\
{\
const int16x8_t p1 = vaddq_s16(vmull_s8(vget_low_s8 (q2bytes.val[0]), vget_low_s8 (q8bytes.val[0])),\
vmull_s8(vget_high_s8(q2bytes.val[0]), vget_high_s8(q8bytes.val[0])));\
const int16x8_t p2 = vaddq_s16(vmull_s8(vget_low_s8 (q2bytes.val[1]), vget_low_s8 (q8bytes.val[1])),\
vmull_s8(vget_high_s8(q2bytes.val[1]), vget_high_s8(q8bytes.val[1])));\
isum += vaddvq_s16(p1) * aux[is+(index)] + vaddvq_s16(p2) * aux[is+1+(index)];\
}
#endif
#define SHIFT_MULTIPLY_ACCUM_WITH_SCALE(shift, index)\
q8bytes = ggml_vld1q_s8_x2(q8); q8 += 32;\
@ -3710,26 +3607,23 @@ void ggml_vec_dot_q2_K_q8_K(const int n, float * restrict s, const void * restri
q2bytes.val[1] = vreinterpretq_s8_u8(vandq_u8(vshrq_n_u8(q2bits.val[1], (shift)), m3));\
MULTIPLY_ACCUM_WITH_SCALE((index));
for (int j = 0; j < QK_K/128; ++j) {
const ggml_uint8x16x2_t q2bits = ggml_vld1q_u8_x2(q2); q2 += 32;
ggml_int8x16x2_t q8bytes = ggml_vld1q_s8_x2(q8); q8 += 32;
q2bytes.val[0] = vreinterpretq_s8_u8(vandq_u8(q2bits.val[0], m3));
q2bytes.val[1] = vreinterpretq_s8_u8(vandq_u8(q2bits.val[1], m3));
MULTIPLY_ACCUM_WITH_SCALE(0);
SHIFT_MULTIPLY_ACCUM_WITH_SCALE(2, 2);
SHIFT_MULTIPLY_ACCUM_WITH_SCALE(4, 4);
SHIFT_MULTIPLY_ACCUM_WITH_SCALE(6, 6);
is += 8;
}
sum += d * isum;
sum += d * isum;
}
*s = sum;
@ -4043,11 +3937,9 @@ void ggml_vec_dot_q2_K_q8_K(const int n, float * restrict s, const void * restri
const int nb = n / QK_K;
#ifdef __ARM_NEON
const uint8x16_t m3 = vdupq_n_u8(0x3);
#if defined(__ARM_FEATURE_DOTPROD)
const int32x4_t vzero = vdupq_n_s32(0);
#endif
const int32x4_t vzero = vdupq_n_s32(0);
ggml_int8x16x4_t q2bytes;
@ -4081,28 +3973,12 @@ void ggml_vec_dot_q2_K_q8_K(const int n, float * restrict s, const void * restri
q2bytes.val[2] = vreinterpretq_s8_u8(vandq_u8(vshrq_n_u8(q2bits, 4), m3));
q2bytes.val[3] = vreinterpretq_s8_u8(vandq_u8(vshrq_n_u8(q2bits, 6), m3));
#if defined(__ARM_FEATURE_DOTPROD)
isum1 += vaddvq_s32(vdotq_s32(vzero, q2bytes.val[0], q8bytes.val[0])) * scales[0];
isum2 += vaddvq_s32(vdotq_s32(vzero, q2bytes.val[1], q8bytes.val[1])) * scales[1];
isum1 += vaddvq_s32(vdotq_s32(vzero, q2bytes.val[2], q8bytes.val[2])) * scales[2];
isum2 += vaddvq_s32(vdotq_s32(vzero, q2bytes.val[3], q8bytes.val[3])) * scales[3];
#else
const int16x8_t p1 = vaddq_s16(vmull_s8(vget_low_s8 (q2bytes.val[0]), vget_low_s8 (q8bytes.val[0])),
vmull_s8(vget_high_s8(q2bytes.val[0]), vget_high_s8(q8bytes.val[0])));
const int16x8_t p2 = vaddq_s16(vmull_s8(vget_low_s8 (q2bytes.val[1]), vget_low_s8 (q8bytes.val[1])),
vmull_s8(vget_high_s8(q2bytes.val[1]), vget_high_s8(q8bytes.val[1])));
isum1 += vaddvq_s16(p1) * scales[0];
isum2 += vaddvq_s16(p2) * scales[1];
const int16x8_t p3 = vaddq_s16(vmull_s8(vget_low_s8 (q2bytes.val[2]), vget_low_s8 (q8bytes.val[2])),
vmull_s8(vget_high_s8(q2bytes.val[2]), vget_high_s8(q8bytes.val[2])));
const int16x8_t p4 = vaddq_s16(vmull_s8(vget_low_s8 (q2bytes.val[3]), vget_low_s8 (q8bytes.val[3])),
vmull_s8(vget_high_s8(q2bytes.val[3]), vget_high_s8(q8bytes.val[3])));
isum1 += vaddvq_s16(p3) * scales[2];
isum2 += vaddvq_s16(p4) * scales[3];
#endif
sum += d * (isum1 + isum2);
}
*s = sum;
@ -4328,9 +4204,7 @@ void ggml_vec_dot_q3_K_q8_K(const int n, float * restrict s, const void * restri
uint32_t utmp[4];
const uint8x16_t m3b = vdupq_n_u8(0x3);
#ifdef __ARM_FEATURE_DOTPROD
const int32x4_t vzero = vdupq_n_s32(0);
#endif
const uint8x16_t m0 = vdupq_n_u8(1);
const uint8x16_t m1 = vshlq_n_u8(m0, 1);
@ -4382,22 +4256,11 @@ void ggml_vec_dot_q3_K_q8_K(const int n, float * restrict s, const void * restri
q3bytes.val[2] = vsubq_s8(vreinterpretq_s8_u8(vandq_u8(vshrq_n_u8(q3bits.val[0], 2), m3b)), vreinterpretq_s8_u8(q3h.val[2]));
q3bytes.val[3] = vsubq_s8(vreinterpretq_s8_u8(vandq_u8(vshrq_n_u8(q3bits.val[1], 2), m3b)), vreinterpretq_s8_u8(q3h.val[3]));
#if defined(__ARM_FEATURE_DOTPROD)
isum += vaddvq_s32(vdotq_s32(vzero, q3bytes.val[0], q8bytes_1.val[0])) * scale[0];
isum += vaddvq_s32(vdotq_s32(vzero, q3bytes.val[1], q8bytes_1.val[1])) * scale[1];
isum += vaddvq_s32(vdotq_s32(vzero, q3bytes.val[2], q8bytes_1.val[2])) * scale[2];
isum += vaddvq_s32(vdotq_s32(vzero, q3bytes.val[3], q8bytes_1.val[3])) * scale[3];
#else
int16x8_t p0 = vaddq_s16(vmull_s8(vget_low_s8 (q3bytes.val[0]), vget_low_s8 (q8bytes_1.val[0])),
vmull_s8(vget_high_s8(q3bytes.val[0]), vget_high_s8(q8bytes_1.val[0])));
int16x8_t p1 = vaddq_s16(vmull_s8(vget_low_s8 (q3bytes.val[1]), vget_low_s8 (q8bytes_1.val[1])),
vmull_s8(vget_high_s8(q3bytes.val[1]), vget_high_s8(q8bytes_1.val[1])));
int16x8_t p2 = vaddq_s16(vmull_s8(vget_low_s8 (q3bytes.val[2]), vget_low_s8 (q8bytes_1.val[2])),
vmull_s8(vget_high_s8(q3bytes.val[2]), vget_high_s8(q8bytes_1.val[2])));
int16x8_t p3 = vaddq_s16(vmull_s8(vget_low_s8 (q3bytes.val[3]), vget_low_s8 (q8bytes_1.val[3])),
vmull_s8(vget_high_s8(q3bytes.val[3]), vget_high_s8(q8bytes_1.val[3])));
isum += vaddvq_s16(p0) * scale[0] + vaddvq_s16(p1) * scale[1] + vaddvq_s16(p2) * scale[2] + vaddvq_s16(p3) * scale[3];
#endif
scale += 4;
q3h.val[0] = vbicq_u8(m2, qhbits.val[0]);
@ -4410,22 +4273,11 @@ void ggml_vec_dot_q3_K_q8_K(const int n, float * restrict s, const void * restri
q3bytes.val[2] = vsubq_s8(vreinterpretq_s8_u8(vandq_u8(vshrq_n_u8(q3bits.val[0], 6), m3b)), vreinterpretq_s8_u8(q3h.val[2]));
q3bytes.val[3] = vsubq_s8(vreinterpretq_s8_u8(vandq_u8(vshrq_n_u8(q3bits.val[1], 6), m3b)), vreinterpretq_s8_u8(q3h.val[3]));
#if defined(__ARM_FEATURE_DOTPROD)
isum += vaddvq_s32(vdotq_s32(vzero, q3bytes.val[0], q8bytes_2.val[0])) * scale[0];
isum += vaddvq_s32(vdotq_s32(vzero, q3bytes.val[1], q8bytes_2.val[1])) * scale[1];
isum += vaddvq_s32(vdotq_s32(vzero, q3bytes.val[2], q8bytes_2.val[2])) * scale[2];
isum += vaddvq_s32(vdotq_s32(vzero, q3bytes.val[3], q8bytes_2.val[3])) * scale[3];
#else
p0 = vaddq_s16(vmull_s8(vget_low_s8 (q3bytes.val[0]), vget_low_s8 (q8bytes_2.val[0])),
vmull_s8(vget_high_s8(q3bytes.val[0]), vget_high_s8(q8bytes_2.val[0])));
p1 = vaddq_s16(vmull_s8(vget_low_s8 (q3bytes.val[1]), vget_low_s8 (q8bytes_2.val[1])),
vmull_s8(vget_high_s8(q3bytes.val[1]), vget_high_s8(q8bytes_2.val[1])));
p2 = vaddq_s16(vmull_s8(vget_low_s8 (q3bytes.val[2]), vget_low_s8 (q8bytes_2.val[2])),
vmull_s8(vget_high_s8(q3bytes.val[2]), vget_high_s8(q8bytes_2.val[2])));
p3 = vaddq_s16(vmull_s8(vget_low_s8 (q3bytes.val[3]), vget_low_s8 (q8bytes_2.val[3])),
vmull_s8(vget_high_s8(q3bytes.val[3]), vget_high_s8(q8bytes_2.val[3])));
isum += vaddvq_s16(p0) * scale[0] + vaddvq_s16(p1) * scale[1] + vaddvq_s16(p2) * scale[2] + vaddvq_s16(p3) * scale[3];
#endif
scale += 4;
if (j == 0) {
@ -4757,7 +4609,7 @@ void ggml_vec_dot_q3_K_q8_K(const int n, float * restrict s, const void * restri
vl = 16;
// retreive lane to multiply with scale
// retrieve lane to multiply with scale
vint32m2_t aux0_0 = __riscv_vwmul_vx_i32m2(__riscv_vget_v_i16m2_i16m1(a0, 0), (scale[0]), vl);
vint32m2_t aux0_1 = __riscv_vwmul_vx_i32m2(__riscv_vget_v_i16m2_i16m1(a0, 1), (scale[1]), vl);
vint32m2_t aux1_0 = __riscv_vwmul_vx_i32m2(__riscv_vget_v_i16m2_i16m1(a1, 0), (scale[2]), vl);
@ -4864,10 +4716,7 @@ void ggml_vec_dot_q3_K_q8_K(const int n, float * restrict s, const void * restri
const int nb = n / QK_K;
#ifdef __ARM_NEON
#ifdef __ARM_FEATURE_DOTPROD
const int32x4_t vzero = vdupq_n_s32(0);
#endif
const int32x4_t vzero = vdupq_n_s32(0);
const uint8x16_t m3b = vdupq_n_u8(0x3);
const uint8x16_t mh = vdupq_n_u8(4);
@ -4908,22 +4757,10 @@ void ggml_vec_dot_q3_K_q8_K(const int n, float * restrict s, const void * restri
q3bytes.val[2] = vreinterpretq_s8_u8(vorrq_u8(vandq_u8(vshrq_n_u8(q3bits, 4), m3b), q3h.val[2]));
q3bytes.val[3] = vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(q3bits, 6), q3h.val[3]));
#if defined(__ARM_FEATURE_DOTPROD)
isum += vaddvq_s32(vdotq_s32(vzero, q3bytes.val[0], q8bytes.val[0])) * scales[0];
isum += vaddvq_s32(vdotq_s32(vzero, q3bytes.val[1], q8bytes.val[1])) * scales[2];
isum += vaddvq_s32(vdotq_s32(vzero, q3bytes.val[2], q8bytes.val[2])) * scales[1];
isum += vaddvq_s32(vdotq_s32(vzero, q3bytes.val[3], q8bytes.val[3])) * scales[3];
#else
const int16x8_t p0 = vaddq_s16(vmull_s8(vget_low_s8 (q3bytes.val[0]), vget_low_s8 (q8bytes.val[0])),
vmull_s8(vget_high_s8(q3bytes.val[0]), vget_high_s8(q8bytes.val[0])));
const int16x8_t p1 = vaddq_s16(vmull_s8(vget_low_s8 (q3bytes.val[1]), vget_low_s8 (q8bytes.val[1])),
vmull_s8(vget_high_s8(q3bytes.val[1]), vget_high_s8(q8bytes.val[1])));
const int16x8_t p2 = vaddq_s16(vmull_s8(vget_low_s8 (q3bytes.val[2]), vget_low_s8 (q8bytes.val[2])),
vmull_s8(vget_high_s8(q3bytes.val[2]), vget_high_s8(q8bytes.val[2])));
const int16x8_t p3 = vaddq_s16(vmull_s8(vget_low_s8 (q3bytes.val[3]), vget_low_s8 (q8bytes.val[3])),
vmull_s8(vget_high_s8(q3bytes.val[3]), vget_high_s8(q8bytes.val[3])));
isum += vaddvq_s16(p0) * scales[0] + vaddvq_s16(p1) * scales[2] + vaddvq_s16(p2) * scales[1] + vaddvq_s16(p3) * scales[3];
#endif
sum += d * isum;
@ -5228,11 +5065,8 @@ void ggml_vec_dot_q4_K_q8_K(const int n, float * restrict s, const void * restri
uint32_t utmp[4];
#ifdef __ARM_NEON
const uint8x16_t m4b = vdupq_n_u8(0xf);
#ifdef __ARM_FEATURE_DOTPROD
const int32x4_t mzero = vdupq_n_s32(0);
#endif
ggml_int8x16x2_t q4bytes;
ggml_int8x16x2_t q8bytes;
@ -5269,10 +5103,8 @@ void ggml_vec_dot_q4_K_q8_K(const int n, float * restrict s, const void * restri
int32_t sumi2 = 0;
for (int j = 0; j < QK_K/64; ++j) {
const ggml_uint8x16x2_t q4bits = ggml_vld1q_u8_x2(q4); q4 += 32;
#ifdef __ARM_FEATURE_DOTPROD
q8bytes = ggml_vld1q_s8_x2(q8); q8 += 32;
q4bytes.val[0] = vreinterpretq_s8_u8(vandq_u8 (q4bits.val[0], m4b));
q4bytes.val[1] = vreinterpretq_s8_u8(vandq_u8 (q4bits.val[1], m4b));
@ -5287,26 +5119,6 @@ void ggml_vec_dot_q4_K_q8_K(const int n, float * restrict s, const void * restri
const int32x4_t p2 = vdotq_s32(vdotq_s32(mzero, q4bytes.val[0], q8bytes.val[0]), q4bytes.val[1], q8bytes.val[1]);
sumi2 += vaddvq_s32(p2) * scales[2*j+1];
#else
q8bytes = ggml_vld1q_s8_x2(q8); q8 += 32;
q4bytes.val[0] = vreinterpretq_s8_u8(vandq_u8 (q4bits.val[0], m4b));
q4bytes.val[1] = vreinterpretq_s8_u8(vandq_u8 (q4bits.val[1], m4b));
const int16x8_t p0 = vaddq_s16(vmull_s8(vget_low_s8 (q4bytes.val[0]), vget_low_s8 (q8bytes.val[0])),
vmull_s8(vget_high_s8(q4bytes.val[0]), vget_high_s8(q8bytes.val[0])));
const int16x8_t p1 = vaddq_s16(vmull_s8(vget_low_s8 (q4bytes.val[1]), vget_low_s8 (q8bytes.val[1])),
vmull_s8(vget_high_s8(q4bytes.val[1]), vget_high_s8(q8bytes.val[1])));
sumi1 += vaddvq_s16(vaddq_s16(p0, p1)) * scales[2*j+0];
q8bytes = ggml_vld1q_s8_x2(q8); q8 += 32;
q4bytes.val[0] = vreinterpretq_s8_u8(vshrq_n_u8(q4bits.val[0], 4));
q4bytes.val[1] = vreinterpretq_s8_u8(vshrq_n_u8(q4bits.val[1], 4));
const int16x8_t p2 = vaddq_s16(vmull_s8(vget_low_s8 (q4bytes.val[0]), vget_low_s8 (q8bytes.val[0])),
vmull_s8(vget_high_s8(q4bytes.val[0]), vget_high_s8(q8bytes.val[0])));
const int16x8_t p3 = vaddq_s16(vmull_s8(vget_low_s8 (q4bytes.val[1]), vget_low_s8 (q8bytes.val[1])),
vmull_s8(vget_high_s8(q4bytes.val[1]), vget_high_s8(q8bytes.val[1])));
sumi2 += vaddvq_s16(vaddq_s16(p2, p3)) * scales[2*j+1];
#endif
}
sumf += d * (sumi1 + sumi2);
@ -5603,12 +5415,9 @@ void ggml_vec_dot_q4_K_q8_K(const int n, float * restrict s, const void * restri
const int nb = n / QK_K;
#ifdef __ARM_NEON
const uint8x16_t m4b = vdupq_n_u8(0xf);
#ifdef __ARM_FEATURE_DOTPROD
const int32x4_t mzero = vdupq_n_s32(0);
#endif
float sumf = 0;
@ -5636,7 +5445,6 @@ void ggml_vec_dot_q4_K_q8_K(const int n, float * restrict s, const void * restri
const ggml_uint8x16x2_t q4bits = ggml_vld1q_u8_x2(q4);
#ifdef __ARM_FEATURE_DOTPROD
q8bytes = ggml_vld1q_s8_x4(q8);
q4bytes.val[0] = vreinterpretq_s8_u8(vandq_u8 (q4bits.val[0], m4b));
q4bytes.val[1] = vreinterpretq_s8_u8(vandq_u8 (q4bits.val[1], m4b));
@ -5650,27 +5458,7 @@ void ggml_vec_dot_q4_K_q8_K(const int n, float * restrict s, const void * restri
const int32x4_t p2 = vdotq_s32(vdotq_s32(mzero, q4bytes.val[0], q8bytes.val[2]), q4bytes.val[1], q8bytes.val[3]);
const int32_t sumi2 = vaddvq_s32(p2) * scales[1];
#else
q8bytes = ggml_vld1q_s8_x4(q8);
q4bytes.val[0] = vreinterpretq_s8_u8(vandq_u8 (q4bits.val[0], m4b));
q4bytes.val[1] = vreinterpretq_s8_u8(vandq_u8 (q4bits.val[1], m4b));
const int16x8_t p0 = vaddq_s16(vmull_s8(vget_low_s8 (q4bytes.val[0]), vget_low_s8 (q8bytes.val[0])),
vmull_s8(vget_high_s8(q4bytes.val[0]), vget_high_s8(q8bytes.val[0])));
const int16x8_t p1 = vaddq_s16(vmull_s8(vget_low_s8 (q4bytes.val[1]), vget_low_s8 (q8bytes.val[1])),
vmull_s8(vget_high_s8(q4bytes.val[1]), vget_high_s8(q8bytes.val[1])));
int32_t sumi1 = vaddvq_s16(vaddq_s16(p0, p1)) * scales[0];
q4bytes.val[0] = vreinterpretq_s8_u8(vshrq_n_u8(q4bits.val[0], 4));
q4bytes.val[1] = vreinterpretq_s8_u8(vshrq_n_u8(q4bits.val[1], 4));
const int16x8_t p2 = vaddq_s16(vmull_s8(vget_low_s8 (q4bytes.val[0]), vget_low_s8 (q8bytes.val[2])),
vmull_s8(vget_high_s8(q4bytes.val[0]), vget_high_s8(q8bytes.val[2])));
const int16x8_t p3 = vaddq_s16(vmull_s8(vget_low_s8 (q4bytes.val[1]), vget_low_s8 (q8bytes.val[3])),
vmull_s8(vget_high_s8(q4bytes.val[1]), vget_high_s8(q8bytes.val[3])));
int32_t sumi2 = vaddvq_s16(vaddq_s16(p2, p3)) * scales[1];
#endif
sumf += d * (sumi1 + sumi2);
}
*s = sumf - sum_mins;
@ -5875,15 +5663,11 @@ void ggml_vec_dot_q5_K_q8_K(const int n, float * restrict s, const void * restri
uint32_t utmp[4];
#ifdef __ARM_NEON
const uint8x16_t m4b = vdupq_n_u8(0xf);
const uint8x16_t mone = vdupq_n_u8(1);
const uint8x16_t mtwo = vdupq_n_u8(2);
#if defined(__ARM_FEATURE_DOTPROD)
const int32x4_t mzero = vdupq_n_s32(0);
#endif
ggml_int8x16x4_t q5bytes;
@ -5938,28 +5722,11 @@ void ggml_vec_dot_q5_K_q8_K(const int n, float * restrict s, const void * restri
q5bytes.val[2] = vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(q5bits.val[0], 4), q5h.val[2]));
q5bytes.val[3] = vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(q5bits.val[1], 4), q5h.val[3]));
#if defined(__ARM_FEATURE_DOTPROD)
sumi += vaddvq_s32(vdotq_s32(vdotq_s32(mzero, q5bytes.val[0], q8bytes.val[0]), q5bytes.val[1], q8bytes.val[1])) * *scales++;
sumi += vaddvq_s32(vdotq_s32(vdotq_s32(mzero, q5bytes.val[2], q8bytes.val[2]), q5bytes.val[3], q8bytes.val[3])) * *scales++;
#else
const int16x8_t p0 = vaddq_s16(vmull_s8(vget_low_s8 (q5bytes.val[0]), vget_low_s8 (q8bytes.val[0])),
vmull_s8(vget_high_s8(q5bytes.val[0]), vget_high_s8(q8bytes.val[0])));
const int16x8_t p1 = vaddq_s16(vmull_s8(vget_low_s8 (q5bytes.val[1]), vget_low_s8 (q8bytes.val[1])),
vmull_s8(vget_high_s8(q5bytes.val[1]), vget_high_s8(q8bytes.val[1])));
sumi += vaddvq_s16(vaddq_s16(p0, p1)) * *scales++;
const int16x8_t p2 = vaddq_s16(vmull_s8(vget_low_s8 (q5bytes.val[2]), vget_low_s8 (q8bytes.val[2])),
vmull_s8(vget_high_s8(q5bytes.val[2]), vget_high_s8(q8bytes.val[2])));
const int16x8_t p3 = vaddq_s16(vmull_s8(vget_low_s8 (q5bytes.val[3]), vget_low_s8 (q8bytes.val[3])),
vmull_s8(vget_high_s8(q5bytes.val[3]), vget_high_s8(q8bytes.val[3])));
sumi += vaddvq_s16(vaddq_s16(p2, p3)) * *scales++;
#endif
}
sumf += d * sumi - dmin * sumi_mins;
}
*s = sumf;
@ -6311,12 +6078,9 @@ void ggml_vec_dot_q5_K_q8_K(const int n, float * restrict s, const void * restri
const int nb = n / QK_K;
#ifdef __ARM_NEON
const uint8x16_t m4b = vdupq_n_u8(0xf);
const uint8x16_t mh = vdupq_n_u8(16);
#if defined(__ARM_FEATURE_DOTPROD)
const int32x4_t mzero = vdupq_n_s32(0);
#endif
ggml_int8x16x4_t q5bytes;
ggml_uint8x16x4_t q5h;
@ -6348,32 +6112,12 @@ void ggml_vec_dot_q5_K_q8_K(const int n, float * restrict s, const void * restri
q5bytes.val[2] = vsubq_s8(vreinterpretq_s8_u8(vshrq_n_u8(q5bits.val[0], 4)), vreinterpretq_s8_u8(q5h.val[2]));
q5bytes.val[3] = vsubq_s8(vreinterpretq_s8_u8(vshrq_n_u8(q5bits.val[1], 4)), vreinterpretq_s8_u8(q5h.val[3]));
#if defined(__ARM_FEATURE_DOTPROD)
int32_t sumi1 = sc[0] * vaddvq_s32(vdotq_s32(mzero, q5bytes.val[0], q8bytes.val[0]));
int32_t sumi2 = sc[1] * vaddvq_s32(vdotq_s32(mzero, q5bytes.val[1], q8bytes.val[1]));
int32_t sumi3 = sc[2] * vaddvq_s32(vdotq_s32(mzero, q5bytes.val[2], q8bytes.val[2]));
int32_t sumi4 = sc[3] * vaddvq_s32(vdotq_s32(mzero, q5bytes.val[3], q8bytes.val[3]));
sumf += d * (sumi1 + sumi2 + sumi3 + sumi4);
#else
const int16x8_t p0 = vaddq_s16(vmull_s8(vget_low_s8 (q5bytes.val[0]), vget_low_s8 (q8bytes.val[0])),
vmull_s8(vget_high_s8(q5bytes.val[0]), vget_high_s8(q8bytes.val[0])));
const int16x8_t p1 = vaddq_s16(vmull_s8(vget_low_s8 (q5bytes.val[1]), vget_low_s8 (q8bytes.val[1])),
vmull_s8(vget_high_s8(q5bytes.val[1]), vget_high_s8(q8bytes.val[1])));
int32_t sumi = sc[0] * vaddvq_s16(p0) + sc[1] * vaddvq_s16(p1);
const int16x8_t p2 = vaddq_s16(vmull_s8(vget_low_s8 (q5bytes.val[2]), vget_low_s8 (q8bytes.val[2])),
vmull_s8(vget_high_s8(q5bytes.val[2]), vget_high_s8(q8bytes.val[2])));
const int16x8_t p3 = vaddq_s16(vmull_s8(vget_low_s8 (q5bytes.val[3]), vget_low_s8 (q8bytes.val[3])),
vmull_s8(vget_high_s8(q5bytes.val[3]), vget_high_s8(q8bytes.val[3])));
sumi += sc[2] * vaddvq_s16(p2) + sc[3] * vaddvq_s16(p3);
sumf += d*sumi;
#endif
}
*s = sumf;
@ -6600,13 +6344,10 @@ void ggml_vec_dot_q6_K_q8_K(const int n, float * restrict s, const void * restri
const int nb = n / QK_K;
#ifdef __ARM_NEON
float sum = 0;
const uint8x16_t m4b = vdupq_n_u8(0xF);
#if defined(__ARM_FEATURE_DOTPROD)
const int32x4_t vzero = vdupq_n_s32(0);
#endif
//const int8x16_t m32s = vdupq_n_s8(32);
const uint8x16_t mone = vdupq_n_u8(3);
@ -6626,7 +6367,7 @@ void ggml_vec_dot_q6_K_q8_K(const int n, float * restrict s, const void * restri
const ggml_int16x8x2_t q8sums = ggml_vld1q_s16_x2(y[i].bsums);
const int8x16_t scales = vld1q_s8(scale);
const ggml_int16x8x2_t q6scales = {vmovl_s8(vget_low_s8(scales)), vmovl_s8(vget_high_s8(scales))};
const ggml_int16x8x2_t q6scales = {{vmovl_s8(vget_low_s8(scales)), vmovl_s8(vget_high_s8(scales))}};
const int32x4_t prod = vaddq_s32(vaddq_s32(vmull_s16(vget_low_s16 (q8sums.val[0]), vget_low_s16 (q6scales.val[0])),
vmull_s16(vget_high_s16(q8sums.val[0]), vget_high_s16(q6scales.val[0]))),
@ -6658,31 +6399,13 @@ void ggml_vec_dot_q6_K_q8_K(const int n, float * restrict s, const void * restri
q6bytes.val[2] = vreinterpretq_s8_u8(vorrq_u8(vandq_u8(q6bits.val[2], m4b), q6h.val[2]));
q6bytes.val[3] = vreinterpretq_s8_u8(vorrq_u8(vandq_u8(q6bits.val[3], m4b), q6h.val[3]));
#if defined(__ARM_FEATURE_DOTPROD)
isum += vaddvq_s32(vdotq_s32(vzero, q6bytes.val[0], q8bytes.val[0])) * scale[0] +
vaddvq_s32(vdotq_s32(vzero, q6bytes.val[1], q8bytes.val[1])) * scale[1] +
vaddvq_s32(vdotq_s32(vzero, q6bytes.val[2], q8bytes.val[2])) * scale[2] +
vaddvq_s32(vdotq_s32(vzero, q6bytes.val[3], q8bytes.val[3])) * scale[3];
scale += 4;
#else
int16x8_t p0 = vaddq_s16(vmull_s8(vget_low_s8 (q6bytes.val[0]), vget_low_s8 (q8bytes.val[0])),
vmull_s8(vget_high_s8(q6bytes.val[0]), vget_high_s8(q8bytes.val[0])));
int16x8_t p1 = vaddq_s16(vmull_s8(vget_low_s8 (q6bytes.val[1]), vget_low_s8 (q8bytes.val[1])),
vmull_s8(vget_high_s8(q6bytes.val[1]), vget_high_s8(q8bytes.val[1])));
isum += vaddvq_s16(p0) * scale[0] + vaddvq_s16(p1) * scale[1];
scale += 2;
int16x8_t p2 = vaddq_s16(vmull_s8(vget_low_s8 (q6bytes.val[2]), vget_low_s8 (q8bytes.val[2])),
vmull_s8(vget_high_s8(q6bytes.val[2]), vget_high_s8(q8bytes.val[2])));
int16x8_t p3 = vaddq_s16(vmull_s8(vget_low_s8 (q6bytes.val[3]), vget_low_s8 (q8bytes.val[3])),
vmull_s8(vget_high_s8(q6bytes.val[3]), vget_high_s8(q8bytes.val[3])));
isum += vaddvq_s16(p2) * scale[0] + vaddvq_s16(p3) * scale[1];
scale += 2;
#endif
q8bytes = ggml_vld1q_s8_x4(q8); q8 += 64;
shifted = vshrq_n_u8(qhbits.val[0], 4);
@ -6703,34 +6426,11 @@ void ggml_vec_dot_q6_K_q8_K(const int n, float * restrict s, const void * restri
q6bytes.val[2] = vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(q6bits.val[2], 4), q6h.val[2]));
q6bytes.val[3] = vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(q6bits.val[3], 4), q6h.val[3]));
#if defined(__ARM_FEATURE_DOTPROD)
isum += vaddvq_s32(vdotq_s32(vzero, q6bytes.val[0], q8bytes.val[0])) * scale[0] +
vaddvq_s32(vdotq_s32(vzero, q6bytes.val[1], q8bytes.val[1])) * scale[1] +
vaddvq_s32(vdotq_s32(vzero, q6bytes.val[2], q8bytes.val[2])) * scale[2] +
vaddvq_s32(vdotq_s32(vzero, q6bytes.val[3], q8bytes.val[3])) * scale[3];
scale += 4;
//for (int l = 0; l < 4; ++l) {
// const int32x4_t p = vdotq_s32(vzero, q6bytes.val[l], q8bytes.val[l]);
// isum += vaddvq_s32(p) * *scale++;
//}
#else
p0 = vaddq_s16(vmull_s8(vget_low_s8 (q6bytes.val[0]), vget_low_s8 (q8bytes.val[0])),
vmull_s8(vget_high_s8(q6bytes.val[0]), vget_high_s8(q8bytes.val[0])));
p1 = vaddq_s16(vmull_s8(vget_low_s8 (q6bytes.val[1]), vget_low_s8 (q8bytes.val[1])),
vmull_s8(vget_high_s8(q6bytes.val[1]), vget_high_s8(q8bytes.val[1])));
isum += vaddvq_s16(p0) * scale[0] + vaddvq_s16(p1) * scale[1];
scale += 2;
p2 = vaddq_s16(vmull_s8(vget_low_s8 (q6bytes.val[2]), vget_low_s8 (q8bytes.val[2])),
vmull_s8(vget_high_s8(q6bytes.val[2]), vget_high_s8(q8bytes.val[2])));
p3 = vaddq_s16(vmull_s8(vget_low_s8 (q6bytes.val[3]), vget_low_s8 (q8bytes.val[3])),
vmull_s8(vget_high_s8(q6bytes.val[3]), vget_high_s8(q8bytes.val[3])));
isum += vaddvq_s16(p2) * scale[0] + vaddvq_s16(p3) * scale[1];
scale += 2;
#endif
}
//sum += isum * d_all * y[i].d;
sum += d_all * y[i].d * (isum - 32 * isum_mins);
@ -7076,14 +6776,11 @@ void ggml_vec_dot_q6_K_q8_K(const int n, float * restrict s, const void * restri
const int nb = n / QK_K;
#ifdef __ARM_NEON
float sum = 0;
const uint8x16_t m4b = vdupq_n_u8(0xF);
const int8x16_t m32s = vdupq_n_s8(32);
#if defined(__ARM_FEATURE_DOTPROD)
const int32x4_t vzero = vdupq_n_s32(0);
#endif
const uint8x16_t mone = vdupq_n_u8(3);
@ -7119,26 +6816,10 @@ void ggml_vec_dot_q6_K_q8_K(const int n, float * restrict s, const void * restri
q6bytes.val[2] = vsubq_s8(vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(q6bits.val[0], 4), q6h.val[2])), m32s);
q6bytes.val[3] = vsubq_s8(vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(q6bits.val[1], 4), q6h.val[3])), m32s);
#if defined(__ARM_FEATURE_DOTPROD)
isum += vaddvq_s32(vdotq_s32(vzero, q6bytes.val[0], q8bytes.val[0])) * scale[0] +
vaddvq_s32(vdotq_s32(vzero, q6bytes.val[1], q8bytes.val[1])) * scale[1] +
vaddvq_s32(vdotq_s32(vzero, q6bytes.val[2], q8bytes.val[2])) * scale[2] +
vaddvq_s32(vdotq_s32(vzero, q6bytes.val[3], q8bytes.val[3])) * scale[3];
#else
int16x8_t p0 = vaddq_s16(vmull_s8(vget_low_s8 (q6bytes.val[0]), vget_low_s8 (q8bytes.val[0])),
vmull_s8(vget_high_s8(q6bytes.val[0]), vget_high_s8(q8bytes.val[0])));
int16x8_t p1 = vaddq_s16(vmull_s8(vget_low_s8 (q6bytes.val[1]), vget_low_s8 (q8bytes.val[1])),
vmull_s8(vget_high_s8(q6bytes.val[1]), vget_high_s8(q8bytes.val[1])));
isum += vaddvq_s16(p0) * scale[0] + vaddvq_s16(p1) * scale[1];
int16x8_t p2 = vaddq_s16(vmull_s8(vget_low_s8 (q6bytes.val[2]), vget_low_s8 (q8bytes.val[2])),
vmull_s8(vget_high_s8(q6bytes.val[2]), vget_high_s8(q8bytes.val[2])));
int16x8_t p3 = vaddq_s16(vmull_s8(vget_low_s8 (q6bytes.val[3]), vget_low_s8 (q8bytes.val[3])),
vmull_s8(vget_high_s8(q6bytes.val[3]), vget_high_s8(q8bytes.val[3])));
isum += vaddvq_s16(p2) * scale[2] + vaddvq_s16(p3) * scale[3];
#endif
sum += isum * d_all * y[i].d;

1232
ggml.c
View file

File diff suppressed because it is too large Load diff

145
ggml.h
View file

@ -215,9 +215,9 @@
#define GGML_QNT_VERSION_FACTOR 1000 // do not change this
#define GGML_MAX_DIMS 4
#define GGML_MAX_PARAMS 1024
#define GGML_MAX_PARAMS 2048
#define GGML_MAX_CONTEXTS 64
#define GGML_MAX_SRC 6
#define GGML_MAX_SRC 10
#define GGML_MAX_NAME 64
#define GGML_MAX_OP_PARAMS 64
#define GGML_DEFAULT_N_THREADS 4
@ -244,11 +244,10 @@
#define GGML_ASSERT(x) \
do { \
if (!(x)) { \
fprintf(stderr, "GGML_ASSERT: %s:%d: %s\n", __FILE__, __LINE__, #x); \
fflush(stderr); \
fflush(stdout); \
fprintf(stderr, "GGML_ASSERT: %s:%d: %s\n", __FILE__, __LINE__, #x); \
ggml_print_backtrace(); \
exit(1); \
abort(); \
} \
} while (0)
@ -256,6 +255,8 @@
#define GGML_UNREACHABLE() GGML_ASSERT(!"statement should not be reached")
#elif defined(__GNUC__)
#define GGML_UNREACHABLE() __builtin_unreachable()
#elif defined(_MSC_VER)
#define GGML_UNREACHABLE() __assume(0)
#else
#define GGML_UNREACHABLE() ((void) 0)
#endif
@ -284,13 +285,27 @@
const type prefix##3 = (pointer)->array[3]; \
GGML_UNUSED(prefix##3);
#define GGML_TENSOR_UNARY_OP_LOCALS \
GGML_TENSOR_LOCALS(int64_t, ne0, src0, ne) \
GGML_TENSOR_LOCALS(size_t, nb0, src0, nb) \
GGML_TENSOR_LOCALS(int64_t, ne, dst, ne) \
GGML_TENSOR_LOCALS(size_t, nb, dst, nb)
#define GGML_TENSOR_BINARY_OP_LOCALS \
GGML_TENSOR_LOCALS(int64_t, ne0, src0, ne) \
GGML_TENSOR_LOCALS(size_t, nb0, src0, nb) \
GGML_TENSOR_LOCALS(int64_t, ne1, src1, ne) \
GGML_TENSOR_LOCALS(size_t, nb1, src1, nb) \
GGML_TENSOR_LOCALS(int64_t, ne, dst, ne) \
GGML_TENSOR_LOCALS(size_t, nb, dst, nb)
#ifdef __cplusplus
extern "C" {
#endif
#if defined(__ARM_NEON) && defined(__CUDACC__)
typedef half ggml_fp16_t;
#elif defined(__ARM_NEON)
#elif defined(__ARM_NEON) && !defined(_MSC_VER)
typedef __fp16 ggml_fp16_t;
#else
typedef uint16_t ggml_fp16_t;
@ -330,6 +345,12 @@ extern "C" {
GGML_TYPE_COUNT,
};
// precision
enum ggml_prec {
GGML_PREC_DEFAULT,
GGML_PREC_F32,
};
enum ggml_backend_type {
GGML_BACKEND_CPU = 0,
GGML_BACKEND_GPU = 10,
@ -382,6 +403,7 @@ extern "C" {
GGML_OP_GROUP_NORM,
GGML_OP_MUL_MAT,
GGML_OP_MUL_MAT_ID,
GGML_OP_OUT_PROD,
GGML_OP_SCALE,
@ -408,8 +430,10 @@ extern "C" {
GGML_OP_CONV_TRANSPOSE_2D,
GGML_OP_POOL_1D,
GGML_OP_POOL_2D,
GGML_OP_UPSCALE, // nearest interpolate
GGML_OP_PAD,
GGML_OP_ARGSORT,
GGML_OP_LEAKY_RELU,
GGML_OP_FLASH_ATTN,
GGML_OP_FLASH_FF,
@ -449,7 +473,8 @@ extern "C" {
GGML_UNARY_OP_GELU,
GGML_UNARY_OP_GELU_QUICK,
GGML_UNARY_OP_SILU,
GGML_UNARY_OP_LEAKY
GGML_UNARY_OP_COUNT,
};
enum ggml_object_type {
@ -461,7 +486,8 @@ extern "C" {
enum ggml_log_level {
GGML_LOG_LEVEL_ERROR = 2,
GGML_LOG_LEVEL_WARN = 3,
GGML_LOG_LEVEL_INFO = 4
GGML_LOG_LEVEL_INFO = 4,
GGML_LOG_LEVEL_DEBUG = 5
};
// ggml object
@ -485,7 +511,6 @@ extern "C" {
struct ggml_backend_buffer * buffer;
int n_dims;
int64_t ne[GGML_MAX_DIMS]; // number of elements
size_t nb[GGML_MAX_DIMS]; // stride in bytes:
// nb[0] = ggml_type_size(type)
@ -517,7 +542,7 @@ extern "C" {
void * extra; // extra things e.g. for ggml-cuda.cu
char padding[12];
char padding[8];
};
static const size_t GGML_TENSOR_SIZE = sizeof(struct ggml_tensor);
@ -622,16 +647,22 @@ extern "C" {
GGML_API int64_t ggml_nrows (const struct ggml_tensor * tensor);
GGML_API size_t ggml_nbytes (const struct ggml_tensor * tensor);
GGML_API size_t ggml_nbytes_pad (const struct ggml_tensor * tensor); // same as ggml_nbytes() but padded to GGML_MEM_ALIGN
GGML_API size_t ggml_nbytes_split(const struct ggml_tensor * tensor, int nrows_split);
GGML_API int ggml_blck_size (enum ggml_type type);
GGML_API size_t ggml_type_size (enum ggml_type type); // size in bytes for all elements in a block
GGML_API float ggml_type_sizef(enum ggml_type type); // ggml_type_size()/ggml_blck_size() as float
GGML_API int ggml_blck_size(enum ggml_type type);
GGML_API size_t ggml_type_size(enum ggml_type type); // size in bytes for all elements in a block
GGML_API size_t ggml_row_size (enum ggml_type type, int64_t ne); // size in bytes for all elements in a row
GGML_DEPRECATED(
GGML_API double ggml_type_sizef(enum ggml_type type), // ggml_type_size()/ggml_blck_size() as float
"use ggml_row_size() instead");
GGML_API const char * ggml_type_name(enum ggml_type type);
GGML_API const char * ggml_op_name (enum ggml_op op);
GGML_API const char * ggml_op_symbol(enum ggml_op op);
GGML_API const char * ggml_unary_op_name(enum ggml_unary_op op);
GGML_API const char * ggml_op_desc(const struct ggml_tensor * t); // unary or op name
GGML_API size_t ggml_element_size(const struct ggml_tensor * tensor);
GGML_API bool ggml_is_quantized(enum ggml_type type);
@ -642,6 +673,11 @@ extern "C" {
GGML_API bool ggml_is_transposed(const struct ggml_tensor * tensor);
GGML_API bool ggml_is_contiguous(const struct ggml_tensor * tensor);
GGML_API bool ggml_is_permuted (const struct ggml_tensor * tensor);
GGML_API bool ggml_is_scalar (const struct ggml_tensor * tensor);
GGML_API bool ggml_is_vector (const struct ggml_tensor * tensor);
GGML_API bool ggml_is_matrix (const struct ggml_tensor * tensor);
GGML_API bool ggml_is_3d (const struct ggml_tensor * tensor);
GGML_API int ggml_n_dims (const struct ggml_tensor * tensor); // returns 1 for scalars
GGML_API bool ggml_are_same_shape(const struct ggml_tensor * t0, const struct ggml_tensor * t1);
@ -702,8 +738,8 @@ extern "C" {
GGML_API struct ggml_tensor * ggml_view_tensor(struct ggml_context * ctx, struct ggml_tensor * src);
// Context tensor enumeration and lookup
GGML_API struct ggml_tensor * ggml_get_first_tensor(struct ggml_context * ctx);
GGML_API struct ggml_tensor * ggml_get_next_tensor (struct ggml_context * ctx, struct ggml_tensor * tensor);
GGML_API struct ggml_tensor * ggml_get_first_tensor(const struct ggml_context * ctx);
GGML_API struct ggml_tensor * ggml_get_next_tensor (const struct ggml_context * ctx, struct ggml_tensor * tensor);
GGML_API struct ggml_tensor * ggml_get_tensor(struct ggml_context * ctx, const char * name);
GGML_API struct ggml_tensor * ggml_set_zero(struct ggml_tensor * tensor);
@ -774,6 +810,9 @@ extern "C" {
struct ggml_tensor * a,
struct ggml_tensor * b);
// dst = a
// view(dst, nb1, nb2, nb3, offset) += b
// return dst
GGML_API struct ggml_tensor * ggml_acc(
struct ggml_context * ctx,
struct ggml_tensor * a,
@ -938,15 +977,14 @@ extern "C" {
struct ggml_context * ctx,
struct ggml_tensor * a);
GGML_API struct ggml_tensor * ggml_leaky(
GGML_API struct ggml_tensor * ggml_leaky_relu(
struct ggml_context * ctx,
struct ggml_tensor * a);
struct ggml_tensor * a, float negative_slope, bool inplace);
GGML_API struct ggml_tensor * ggml_relu_inplace(
struct ggml_context * ctx,
struct ggml_tensor * a);
// TODO: double-check this computation is correct
GGML_API struct ggml_tensor * ggml_gelu(
struct ggml_context * ctx,
struct ggml_tensor * a);
@ -1028,6 +1066,22 @@ extern "C" {
struct ggml_tensor * a,
struct ggml_tensor * b);
// change the precision of a matrix multiplication
// set to GGML_PREC_F32 for higher precision (useful for phi-2)
GGML_API void ggml_mul_mat_set_prec(
struct ggml_tensor * a,
enum ggml_prec prec);
// indirect matrix multiplication
// ggml_mul_mat_id(ctx, as, ids, id, b) ~= ggml_mul_mat(as[ids[id]], b)
GGML_API struct ggml_tensor * ggml_mul_mat_id(
struct ggml_context * ctx,
struct ggml_tensor * const as[],
int n_as,
struct ggml_tensor * ids,
int id,
struct ggml_tensor * b);
// A: m columns, n rows,
// B: p columns, n rows,
// result is m columns, p rows
@ -1043,13 +1097,13 @@ extern "C" {
GGML_API struct ggml_tensor * ggml_scale(
struct ggml_context * ctx,
struct ggml_tensor * a,
struct ggml_tensor * b);
float s);
// in-place, returns view(a)
GGML_API struct ggml_tensor * ggml_scale_inplace(
struct ggml_context * ctx,
struct ggml_tensor * a,
struct ggml_tensor * b);
float s);
// b -> view(a,offset,nb1,nb2,3), return modified a
GGML_API struct ggml_tensor * ggml_set(
@ -1235,6 +1289,7 @@ extern "C" {
struct ggml_context * ctx,
struct ggml_tensor * a);
// supports 3D: a->ne[2] == b->ne[1]
GGML_API struct ggml_tensor * ggml_get_rows(
struct ggml_context * ctx,
struct ggml_tensor * a,
@ -1283,6 +1338,14 @@ extern "C" {
struct ggml_context * ctx,
struct ggml_tensor * a);
// fused soft_max(a*scale + mask)
// mask is optional
GGML_API struct ggml_tensor * ggml_soft_max_ext(
struct ggml_context * ctx,
struct ggml_tensor * a,
struct ggml_tensor * mask,
float scale);
GGML_API struct ggml_tensor * ggml_soft_max_back(
struct ggml_context * ctx,
struct ggml_tensor * a,
@ -1513,6 +1576,32 @@ extern "C" {
struct ggml_tensor * a,
int scale_factor);
// pad each dimension with zeros: [x, ..., x] -> [x, ..., x, 0, ..., 0]
GGML_API struct ggml_tensor * ggml_pad(
struct ggml_context * ctx,
struct ggml_tensor * a,
int p0,
int p1,
int p2,
int p3);
// sort rows
enum ggml_sort_order {
GGML_SORT_ASC,
GGML_SORT_DESC,
};
GGML_API struct ggml_tensor * ggml_argsort(
struct ggml_context * ctx,
struct ggml_tensor * a,
enum ggml_sort_order order);
// top k elements per row
GGML_API struct ggml_tensor * ggml_top_k(
struct ggml_context * ctx,
struct ggml_tensor * a,
int k);
GGML_API struct ggml_tensor * ggml_flash_attn(
struct ggml_context * ctx,
struct ggml_tensor * q,
@ -1574,7 +1663,6 @@ extern "C" {
int kh);
// used in sam
GGML_API struct ggml_tensor * ggml_add_rel_pos(
struct ggml_context * ctx,
struct ggml_tensor * a,
@ -1749,7 +1837,7 @@ extern "C" {
GGML_API struct ggml_cgraph * ggml_new_graph (struct ggml_context * ctx); // size = GGML_DEFAULT_GRAPH_SIZE, grads = false
GGML_API struct ggml_cgraph * ggml_new_graph_custom (struct ggml_context * ctx, size_t size, bool grads);
GGML_API struct ggml_cgraph * ggml_graph_dup (struct ggml_context * ctx, struct ggml_cgraph * cgraph);
GGML_API struct ggml_cgraph * ggml_graph_view (struct ggml_context * ctx, struct ggml_cgraph * cgraph, int i0, int i1);
GGML_API struct ggml_cgraph ggml_graph_view (struct ggml_cgraph * cgraph, int i0, int i1);
GGML_API void ggml_graph_cpy (struct ggml_cgraph * src, struct ggml_cgraph * dst);
GGML_API void ggml_graph_reset (struct ggml_cgraph * cgraph); // zero grads
GGML_API void ggml_graph_clear (struct ggml_cgraph * cgraph);
@ -2050,10 +2138,11 @@ extern "C" {
GGML_API const void * gguf_get_arr_data(const struct gguf_context * ctx, int key_id);
GGML_API const char * gguf_get_arr_str (const struct gguf_context * ctx, int key_id, int i);
GGML_API int gguf_get_n_tensors (const struct gguf_context * ctx);
GGML_API int gguf_find_tensor (const struct gguf_context * ctx, const char * name);
GGML_API size_t gguf_get_tensor_offset(const struct gguf_context * ctx, int i);
GGML_API char * gguf_get_tensor_name (const struct gguf_context * ctx, int i);
GGML_API int gguf_get_n_tensors (const struct gguf_context * ctx);
GGML_API int gguf_find_tensor (const struct gguf_context * ctx, const char * name);
GGML_API size_t gguf_get_tensor_offset(const struct gguf_context * ctx, int i);
GGML_API char * gguf_get_tensor_name (const struct gguf_context * ctx, int i);
GGML_API enum ggml_type gguf_get_tensor_type (const struct gguf_context * ctx, int i);
// overrides existing values or adds a new one
GGML_API void gguf_set_val_u8 (struct gguf_context * ctx, const char * key, uint8_t val);

4
gguf-py/README.md
View file

@ -3,7 +3,7 @@
This is a Python package for writing binary files in the [GGUF](https://github.com/ggerganov/ggml/pull/302)
(GGML Universal File) format.
See [convert-llama-hf-to-gguf.py](https://github.com/ggerganov/llama.cpp/blob/master/convert-llama-hf-to-gguf.py)
See [convert-llama-hf-to-gguf.py](https://github.com/ggerganov/llama.cpp/blob/master/convert-hf-to-gguf.py)
as an example for its usage.
## Installation
@ -61,7 +61,7 @@ If you want to publish the package manually for any reason, you need to have `tw
pip install build twine
```
Then, folow these steps to release a new version:
Then, follow these steps to release a new version:
1. Bump the version in `pyproject.toml`.
2. Build the package:

69
gguf-py/gguf/constants.py
View file

@ -38,6 +38,8 @@ class Keys:
FEED_FORWARD_LENGTH = "{arch}.feed_forward_length"
USE_PARALLEL_RESIDUAL = "{arch}.use_parallel_residual"
TENSOR_DATA_LAYOUT = "{arch}.tensor_data_layout"
EXPERT_COUNT = "{arch}.expert_count"
EXPERT_USED_COUNT = "{arch}.expert_used_count"
class Attention:
HEAD_COUNT = "{arch}.attention.head_count"
@ -92,6 +94,9 @@ class MODEL_ARCH(IntEnum):
BERT = auto()
BLOOM = auto()
STABLELM = auto()
QWEN = auto()
PHI2 = auto()
PLAMO = auto()
class MODEL_TENSOR(IntEnum):
@ -110,10 +115,15 @@ class MODEL_TENSOR(IntEnum):
ATTN_NORM = auto()
ATTN_NORM_2 = auto()
ATTN_ROT_EMBD = auto()
FFN_GATE_INP = auto()
FFN_NORM = auto()
FFN_GATE = auto()
FFN_DOWN = auto()
FFN_UP = auto()
FFN_NORM = auto()
FFN_ACT = auto()
FFN_GATE_EXP = auto()
FFN_DOWN_EXP = auto()
FFN_UP_EXP = auto()
ATTN_Q_NORM = auto()
ATTN_K_NORM = auto()
@ -132,6 +142,9 @@ MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = {
MODEL_ARCH.BERT: "bert",
MODEL_ARCH.BLOOM: "bloom",
MODEL_ARCH.STABLELM: "stablelm",
MODEL_ARCH.QWEN: "qwen",
MODEL_ARCH.PHI2: "phi2",
MODEL_ARCH.PLAMO: "plamo",
}
TENSOR_NAMES: dict[MODEL_TENSOR, str] = {
@ -152,10 +165,15 @@ TENSOR_NAMES: dict[MODEL_TENSOR, str] = {
MODEL_TENSOR.ATTN_ROT_EMBD: "blk.{bid}.attn_rot_embd",
MODEL_TENSOR.ATTN_Q_NORM: "blk.{bid}.attn_q_norm",
MODEL_TENSOR.ATTN_K_NORM: "blk.{bid}.attn_k_norm",
MODEL_TENSOR.FFN_GATE_INP: "blk.{bid}.ffn_gate_inp",
MODEL_TENSOR.FFN_NORM: "blk.{bid}.ffn_norm",
MODEL_TENSOR.FFN_GATE: "blk.{bid}.ffn_gate",
MODEL_TENSOR.FFN_DOWN: "blk.{bid}.ffn_down",
MODEL_TENSOR.FFN_UP: "blk.{bid}.ffn_up",
MODEL_TENSOR.FFN_ACT: "blk.{bid}.ffn",
MODEL_TENSOR.FFN_GATE_EXP: "blk.{bid}.ffn_gate.{xid}",
MODEL_TENSOR.FFN_DOWN_EXP: "blk.{bid}.ffn_down.{xid}",
MODEL_TENSOR.FFN_UP_EXP: "blk.{bid}.ffn_up.{xid}",
}
MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
@ -170,10 +188,14 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
MODEL_TENSOR.ATTN_V,
MODEL_TENSOR.ATTN_OUT,
MODEL_TENSOR.ATTN_ROT_EMBD,
MODEL_TENSOR.FFN_GATE_INP,
MODEL_TENSOR.FFN_NORM,
MODEL_TENSOR.FFN_GATE,
MODEL_TENSOR.FFN_DOWN,
MODEL_TENSOR.FFN_UP,
MODEL_TENSOR.FFN_GATE_EXP,
MODEL_TENSOR.FFN_DOWN_EXP,
MODEL_TENSOR.FFN_UP_EXP,
],
MODEL_ARCH.GPTNEOX: [
MODEL_TENSOR.TOKEN_EMBD,
@ -249,6 +271,7 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
MODEL_TENSOR.FFN_NORM,
MODEL_TENSOR.FFN_DOWN,
MODEL_TENSOR.FFN_UP,
MODEL_TENSOR.FFN_ACT,
],
MODEL_ARCH.GPTJ: [
MODEL_TENSOR.TOKEN_EMBD,
@ -317,9 +340,49 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
MODEL_TENSOR.FFN_DOWN,
MODEL_TENSOR.FFN_UP,
],
MODEL_ARCH.QWEN: [
MODEL_TENSOR.TOKEN_EMBD,
MODEL_TENSOR.OUTPUT_NORM,
MODEL_TENSOR.OUTPUT,
MODEL_TENSOR.ROPE_FREQS,
MODEL_TENSOR.ATTN_NORM,
MODEL_TENSOR.ATTN_QKV,
MODEL_TENSOR.ATTN_OUT,
MODEL_TENSOR.ATTN_ROT_EMBD,
MODEL_TENSOR.FFN_NORM,
MODEL_TENSOR.FFN_GATE,
MODEL_TENSOR.FFN_DOWN,
MODEL_TENSOR.FFN_UP,
],
MODEL_ARCH.PLAMO: [
MODEL_TENSOR.TOKEN_EMBD,
MODEL_TENSOR.OUTPUT_NORM,
MODEL_TENSOR.OUTPUT,
MODEL_TENSOR.ROPE_FREQS,
MODEL_TENSOR.ATTN_NORM,
MODEL_TENSOR.ATTN_Q,
MODEL_TENSOR.ATTN_K,
MODEL_TENSOR.ATTN_V,
MODEL_TENSOR.ATTN_OUT,
MODEL_TENSOR.ATTN_ROT_EMBD,
MODEL_TENSOR.FFN_GATE,
MODEL_TENSOR.FFN_DOWN,
MODEL_TENSOR.FFN_UP,
],
MODEL_ARCH.GPT2: [
# TODO
],
MODEL_ARCH.PHI2: [
MODEL_TENSOR.TOKEN_EMBD,
MODEL_TENSOR.OUTPUT_NORM,
MODEL_TENSOR.OUTPUT,
MODEL_TENSOR.ATTN_NORM,
MODEL_TENSOR.ATTN_QKV,
MODEL_TENSOR.ATTN_OUT,
MODEL_TENSOR.FFN_NORM,
MODEL_TENSOR.FFN_DOWN,
MODEL_TENSOR.FFN_UP,
]
# TODO
}
@ -336,6 +399,10 @@ MODEL_TENSOR_SKIP: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
MODEL_ARCH.PERSIMMON: [
MODEL_TENSOR.ROPE_FREQS,
],
MODEL_ARCH.QWEN: [
MODEL_TENSOR.ROPE_FREQS,
MODEL_TENSOR.ATTN_ROT_EMBD,
],
}
#

6
gguf-py/gguf/gguf_writer.py
View file

@ -339,6 +339,12 @@ class GGUFWriter:
def add_clamp_kqv(self, value: float) -> None:
self.add_float32(Keys.Attention.CLAMP_KQV.format(arch=self.arch), value)
def add_expert_count(self, count: int) -> None:
self.add_uint32(Keys.LLM.EXPERT_COUNT.format(arch=self.arch), count)
def add_expert_used_count(self, count: int) -> None:
self.add_uint32(Keys.LLM.EXPERT_USED_COUNT.format(arch=self.arch), count)
def add_layer_norm_eps(self, value: float) -> None:
self.add_float32(Keys.Attention.LAYERNORM_EPS.format(arch=self.arch), value)

105
gguf-py/gguf/tensor_mapping.py
View file

@ -10,13 +10,14 @@ class TensorNameMap:
# Token embeddings
MODEL_TENSOR.TOKEN_EMBD: (
"gpt_neox.embed_in", # gptneox
"transformer.wte", # gpt2 gpt-j mpt refact
"transformer.wte", # gpt2 gpt-j mpt refact qwen
"transformer.word_embeddings", # falcon
"word_embeddings", # bloom
"model.embed_tokens", # llama-hf
"tok_embeddings", # llama-pth
"embeddings.word_embeddings", # bert
"language_model.embedding.word_embeddings", # persimmon
"transformer.embd.wte", # phi2
),
# Token type embeddings
@ -38,9 +39,10 @@ class TensorNameMap:
# Output
MODEL_TENSOR.OUTPUT: (
"embed_out", # gptneox
"lm_head", # gpt2 mpt falcon llama-hf baichuan
"lm_head", # gpt2 mpt falcon llama-hf baichuan qwen
"output", # llama-pth bloom
"word_embeddings_for_head", # persimmon
"lm_head.linear", # phi2
),
# Output norm
@ -51,8 +53,9 @@ class TensorNameMap:
"norm", # llama-pth
"embeddings.LayerNorm", # bert
"transformer.norm_f", # mpt
"ln_f", # refact bloom
"ln_f", # refact bloom qwen
"language_model.encoder.final_layernorm", # persimmon
"lm_head.ln", # phi2
),
# Rope frequencies
@ -65,7 +68,7 @@ class TensorNameMap:
# Attention norm
MODEL_TENSOR.ATTN_NORM: (
"gpt_neox.layers.{bid}.input_layernorm", # gptneox
"transformer.h.{bid}.ln_1", # gpt2 gpt-j refact
"transformer.h.{bid}.ln_1", # gpt2 gpt-j refact qwen
"transformer.blocks.{bid}.norm_1", # mpt
"transformer.h.{bid}.input_layernorm", # falcon7b
"h.{bid}.input_layernorm", # bloom
@ -75,6 +78,8 @@ class TensorNameMap:
"encoder.layer.{bid}.attention.output.LayerNorm", # bert
"language_model.encoder.layers.{bid}.input_layernorm", # persimmon
"model.layers.{bid}.ln1", # yi
"transformer.h.{bid}.ln", # phi2
"model.layers.layers.{bid}.norm", # plamo
),
# Attention norm 2
@ -85,41 +90,45 @@ class TensorNameMap:
# Attention query-key-value
MODEL_TENSOR.ATTN_QKV: (
"gpt_neox.layers.{bid}.attention.query_key_value", # gptneox
"transformer.h.{bid}.attn.c_attn", # gpt2
"transformer.h.{bid}.attn.c_attn", # gpt2 qwen
"transformer.blocks.{bid}.attn.Wqkv", # mpt
"transformer.h.{bid}.self_attention.query_key_value", # falcon
"h.{bid}.self_attention.query_key_value", # bloom
"language_model.encoder.layers.{bid}.self_attention.query_key_value", # persimmon
"transformer.h.{bid}.mixer.Wqkv", # phi2
),
# Attention query
MODEL_TENSOR.ATTN_Q: (
"model.layers.{bid}.self_attn.q_proj", # llama-hf
"layers.{bid}.attention.wq", # llama-pth
"encoder.layer.{bid}.attention.self.query", # bert
"transformer.h.{bid}.attn.q_proj", # gpt-j
"model.layers.{bid}.self_attn.q_proj", # llama-hf
"layers.{bid}.attention.wq", # llama-pth
"encoder.layer.{bid}.attention.self.query", # bert
"transformer.h.{bid}.attn.q_proj", # gpt-j
"model.layers.layers.{bid}.self_attn.q_proj", # plamo
),
# Attention key
MODEL_TENSOR.ATTN_K: (
"model.layers.{bid}.self_attn.k_proj", # llama-hf
"layers.{bid}.attention.wk", # llama-pth
"encoder.layer.{bid}.attention.self.key", # bert
"transformer.h.{bid}.attn.k_proj", # gpt-j
"model.layers.{bid}.self_attn.k_proj", # llama-hf
"layers.{bid}.attention.wk", # llama-pth
"encoder.layer.{bid}.attention.self.key", # bert
"transformer.h.{bid}.attn.k_proj", # gpt-j
"model.layers.layers.{bid}.self_attn.k_proj", # plamo
),
# Attention value
MODEL_TENSOR.ATTN_V: (
"model.layers.{bid}.self_attn.v_proj", # llama-hf
"layers.{bid}.attention.wv", # llama-pth
"encoder.layer.{bid}.attention.self.value", # bert
"transformer.h.{bid}.attn.v_proj", # gpt-j
"model.layers.{bid}.self_attn.v_proj", # llama-hf
"layers.{bid}.attention.wv", # llama-pth
"encoder.layer.{bid}.attention.self.value", # bert
"transformer.h.{bid}.attn.v_proj", # gpt-j
"model.layers.layers.{bid}.self_attn.v_proj", # plamo
),
# Attention output
MODEL_TENSOR.ATTN_OUT: (
"gpt_neox.layers.{bid}.attention.dense", # gptneox
"transformer.h.{bid}.attn.c_proj", # gpt2 refact
"transformer.h.{bid}.attn.c_proj", # gpt2 refact qwen
"transformer.blocks.{bid}.attn.out_proj", # mpt
"transformer.h.{bid}.self_attention.dense", # falcon
"h.{bid}.self_attention.dense", # bloom
@ -128,18 +137,21 @@ class TensorNameMap:
"encoder.layer.{bid}.attention.output.dense", # bert
"transformer.h.{bid}.attn.out_proj", # gpt-j
"language_model.encoder.layers.{bid}.self_attention.dense", # persimmon
"transformer.h.{bid}.mixer.out_proj", # phi2
"model.layers.layers.{bid}.self_attn.o_proj", # plamo
),
# Rotary embeddings
MODEL_TENSOR.ATTN_ROT_EMBD: (
"model.layers.{bid}.self_attn.rotary_emb.inv_freq", # llama-hf
"layers.{bid}.attention.inner_attention.rope.freqs", # llama-pth
"model.layers.{bid}.self_attn.rotary_emb.inv_freq", # llama-hf
"layers.{bid}.attention.inner_attention.rope.freqs", # llama-pth
"model.layers.layers.{bid}.self_attn.rotary_emb.inv_freq", # plamo
),
# Feed-forward norm
MODEL_TENSOR.FFN_NORM: (
"gpt_neox.layers.{bid}.post_attention_layernorm", # gptneox
"transformer.h.{bid}.ln_2", # gpt2 refact
"transformer.h.{bid}.ln_2", # gpt2 refact qwen
"h.{bid}.post_attention_layernorm", # bloom
"transformer.blocks.{bid}.norm_2", # mpt
"model.layers.{bid}.post_attention_layernorm", # llama-hf
@ -149,6 +161,11 @@ class TensorNameMap:
"model.layers.{bid}.ln2", # yi
),
MODEL_TENSOR.FFN_GATE_INP: (
"layers.{bid}.feed_forward.gate", # mixtral
"model.layers.{bid}.block_sparse_moe.gate", # mixtral
),
# Feed-forward up
MODEL_TENSOR.FFN_UP: (
"gpt_neox.layers.{bid}.mlp.dense_h_to_4h", # gptneox
@ -161,18 +178,38 @@ class TensorNameMap:
"encoder.layer.{bid}.intermediate.dense", # bert
"transformer.h.{bid}.mlp.fc_in", # gpt-j
"language_model.encoder.layers.{bid}.mlp.dense_h_to_4h", # persimmon
"transformer.h.{bid}.mlp.w1", # qwen
"transformer.h.{bid}.mlp.fc1", # phi2
"model.layers.layers.{bid}.mlp.up_proj", # plamo
),
MODEL_TENSOR.FFN_UP_EXP: (
"layers.{bid}.feed_forward.experts.{xid}.w3", # mixtral
"model.layers.{bid}.block_sparse_moe.experts.{xid}.w3", # mixtral
),
# AWQ-activation gate
MODEL_TENSOR.FFN_ACT: (
"transformer.blocks.{bid}.ffn.act", # mpt
),
# Feed-forward gate
MODEL_TENSOR.FFN_GATE: (
"model.layers.{bid}.mlp.gate_proj", # llama-hf refact
"layers.{bid}.feed_forward.w1", # llama-pth
"model.layers.{bid}.mlp.gate_proj", # llama-hf refact
"layers.{bid}.feed_forward.w1", # llama-pth
"transformer.h.{bid}.mlp.w2", # qwen
"model.layers.layers.{bid}.mlp.gate_proj", # plamo
),
MODEL_TENSOR.FFN_GATE_EXP: (
"layers.{bid}.feed_forward.experts.{xid}.w1", # mixtral
"model.layers.{bid}.block_sparse_moe.experts.{xid}.w1", # mixtral
),
# Feed-forward down
MODEL_TENSOR.FFN_DOWN: (
"gpt_neox.layers.{bid}.mlp.dense_4h_to_h", # gptneox
"transformer.h.{bid}.mlp.c_proj", # gpt2 refact
"transformer.h.{bid}.mlp.c_proj", # gpt2 refact qwen
"transformer.blocks.{bid}.ffn.down_proj", # mpt
"transformer.h.{bid}.mlp.dense_4h_to_h", # falcon
"h.{bid}.mlp.dense_4h_to_h", # bloom
@ -181,6 +218,13 @@ class TensorNameMap:
"encoder.layer.{bid}.output.dense", # bert
"transformer.h.{bid}.mlp.fc_out", # gpt-j
"language_model.encoder.layers.{bid}.mlp.dense_4h_to_h", # persimmon
"transformer.h.{bid}.mlp.fc2", # phi2
"model.layers.layers.{bid}.mlp.down_proj", # plamo
),
MODEL_TENSOR.FFN_DOWN_EXP: (
"layers.{bid}.feed_forward.experts.{xid}.w2", # mixtral
"model.layers.{bid}.block_sparse_moe.experts.{xid}.w2", # mixtral
),
MODEL_TENSOR.ATTN_Q_NORM: (
@ -211,11 +255,14 @@ class TensorNameMap:
for tensor, keys in self.block_mappings_cfg.items():
if tensor not in MODEL_TENSORS[arch]:
continue
tensor_name = TENSOR_NAMES[tensor].format(bid = bid)
self.mapping[tensor_name] = (tensor, tensor_name)
for key in keys:
key = key.format(bid = bid)
self.mapping[key] = (tensor, tensor_name)
# TODO: make this configurable
n_experts = 8
for xid in range(n_experts):
tensor_name = TENSOR_NAMES[tensor].format(bid = bid, xid = xid)
self.mapping[tensor_name] = (tensor, tensor_name)
for key in keys:
key = key.format(bid = bid, xid = xid)
self.mapping[key] = (tensor, tensor_name)
def get_type_and_name(self, key: str, try_suffixes: Sequence[str] = ()) -> tuple[MODEL_TENSOR, str] | None:
result = self.mapping.get(key)

6
gguf-py/gguf/vocab.py
View file

@ -84,7 +84,7 @@ class SpecialVocab:
merges_file = path / 'merges.txt'
if not merges_file.is_file():
return False
with open(merges_file, 'r') as fp:
with open(merges_file, 'r', encoding = 'utf-8') as fp:
first_line = next(fp, '').strip()
if not first_line.startswith('#'):
fp.seek(0)
@ -109,8 +109,10 @@ class SpecialVocab:
return True
def _set_special_token(self, typ: str, tid: Any) -> None:
if not isinstance(tid, int) or tid < 0:
if not isinstance(tid, int):
return
if tid < 0:
raise ValueError(f'invalid value for special token type {typ}: {tid}')
if self.n_vocab is None or tid < self.n_vocab:
if typ in self.special_token_ids:
return

2
gguf-py/pyproject.toml
View file

@ -1,6 +1,6 @@
[tool.poetry]
name = "gguf"
version = "0.6.0"
version = "0.7.0"
description = "Read and write ML models in GGUF for GGML"
authors = ["GGML <ggml@ggml.ai>"]
packages = [

3345
llama.cpp
View file

File diff suppressed because it is too large Load diff

103
llama.h
View file

@ -39,10 +39,11 @@
#define LLAMA_MAX_RNG_STATE (64*1024)
#define LLAMA_FILE_MAGIC_GGLA 0x67676c61u // 'ggla'
#define LLAMA_FILE_MAGIC_GGSN 0x6767736eu // 'ggsn'
#define LLAMA_SESSION_MAGIC LLAMA_FILE_MAGIC_GGSN
#define LLAMA_SESSION_VERSION 2
#define LLAMA_SESSION_VERSION 3
#if defined(GGML_USE_CUBLAS) || defined(GGML_USE_CLBLAST) || defined(GGML_USE_METAL)
// Defined when llama.cpp is compiled with support for offloading model layers to GPU.
@ -128,7 +129,7 @@ extern "C" {
bool sorted;
} llama_token_data_array;
typedef void (*llama_progress_callback)(float progress, void *ctx);
typedef bool (*llama_progress_callback)(float progress, void *ctx);
// Input data for llama_decode
// A llama_batch object can contain input about one or many sequences
@ -160,16 +161,38 @@ extern "C" {
llama_seq_id all_seq_id; // used if seq_id == NULL
} llama_batch;
enum llama_model_kv_override_type {
LLAMA_KV_OVERRIDE_INT,
LLAMA_KV_OVERRIDE_FLOAT,
LLAMA_KV_OVERRIDE_BOOL,
};
struct llama_model_kv_override {
char key[128];
enum llama_model_kv_override_type tag;
union {
int64_t int_value;
double float_value;
bool bool_value;
};
};
struct llama_model_params {
int32_t n_gpu_layers; // number of layers to store in VRAM
int32_t main_gpu; // the GPU that is used for scratch and small tensors
const float * tensor_split; // how to split layers across multiple GPUs (size: LLAMA_MAX_DEVICES)
// called with a progress value between 0 and 1, pass NULL to disable
// Called with a progress value between 0.0 and 1.0. Pass NULL to disable.
// If the provided progress_callback returns true, model loading continues.
// If it returns false, model loading is immediately aborted.
llama_progress_callback progress_callback;
// context pointer passed to the progress callback
void * progress_callback_user_data;
// override key-value pairs of the model meta data
const struct llama_model_kv_override * kv_overrides;
// Keep the booleans together to avoid misalignment during copy-by-value.
bool vocab_only; // only load the vocabulary, no weights
bool use_mmap; // use mmap if possible
@ -187,17 +210,20 @@ extern "C" {
// ref: https://github.com/ggerganov/llama.cpp/pull/2054
float rope_freq_base; // RoPE base frequency, 0 = from model
float rope_freq_scale; // RoPE frequency scaling factor, 0 = from model
float yarn_ext_factor; // YaRN extrapolation mix factor, NaN = from model
float yarn_ext_factor; // YaRN extrapolation mix factor, negative = from model
float yarn_attn_factor; // YaRN magnitude scaling factor
float yarn_beta_fast; // YaRN low correction dim
float yarn_beta_slow; // YaRN high correction dim
uint32_t yarn_orig_ctx; // YaRN original context size
enum ggml_type type_k; // data type for K cache
enum ggml_type type_v; // data type for V cache
// Keep the booleans together to avoid misalignment during copy-by-value.
bool mul_mat_q; // if true, use experimental mul_mat_q kernels (DEPRECATED - always true)
bool f16_kv; // use fp16 for KV cache, fp32 otherwise
bool logits_all; // the llama_eval() call computes all logits, not just the last one
bool embedding; // embedding mode only
bool mul_mat_q; // if true, use experimental mul_mat_q kernels (DEPRECATED - always true)
bool logits_all; // the llama_eval() call computes all logits, not just the last one (DEPRECATED - set llama_batch.logits instead)
bool embedding; // embedding mode only
bool offload_kqv; // whether to offload the KQV ops (including the KV cache) to GPU
};
// model quantization parameters
@ -292,7 +318,9 @@ extern "C" {
LLAMA_API const struct llama_model * llama_get_model(const struct llama_context * ctx);
LLAMA_API int llama_n_ctx (const struct llama_context * ctx);
// TODO: become more consistent with returned int types across the API
LLAMA_API uint32_t llama_n_ctx (const struct llama_context * ctx);
LLAMA_API uint32_t llama_n_batch (const struct llama_context * ctx);
LLAMA_API enum llama_vocab_type llama_vocab_type(const struct llama_model * model);
@ -363,9 +391,60 @@ extern "C" {
// KV cache
//
// Returns the number of tokens in the KV cache
LLAMA_API DEPRECATED(int llama_get_kv_cache_token_count(const struct llama_context * ctx),
"avoid using this, it will be removed in the future, instead - count the tokens in user code");
// Information associated with an individual cell in the KV cache view.
struct llama_kv_cache_view_cell {
// The position for this cell. Takes KV cache shifts into account.
// May be negative if the cell is not populated.
llama_pos pos;
};
// An updateable view of the KV cache.
struct llama_kv_cache_view {
// Number of KV cache cells. This will be the same as the context size.
int32_t n_cells;
// Maximum number of sequences that can exist in a cell. It's not an error
// if there are more sequences in a cell than this value, however they will
// not be visible in the view cells_sequences.
int32_t n_max_seq;
// Number of tokens in the cache. For example, if there are two populated
// cells, the first with 1 sequence id in it and the second with 2 sequence
// ids then you'll have 3 tokens.
int32_t token_count;
// Number of populated cache cells.
int32_t used_cells;
// Maximum contiguous empty slots in the cache.
int32_t max_contiguous;
// Index to the start of the max_contiguous slot range. Can be negative
// when cache is full.
int32_t max_contiguous_idx;
// Information for an individual cell.
struct llama_kv_cache_view_cell * cells;
// The sequences for each cell. There will be n_max_seq items per cell.
llama_seq_id * cells_sequences;
};
// Create an empty KV cache view. (use only for debugging purposes)
LLAMA_API struct llama_kv_cache_view llama_kv_cache_view_init(const struct llama_context * ctx, int32_t n_max_seq);
// Free a KV cache view. (use only for debugging purposes)
LLAMA_API void llama_kv_cache_view_free(struct llama_kv_cache_view * view);
// Update the KV cache view structure with the current state of the KV cache. (use only for debugging purposes)
LLAMA_API void llama_kv_cache_view_update(const struct llama_context * ctx, struct llama_kv_cache_view * view);
// Returns the number of tokens in the KV cache (slow, use only for debug)
// If a KV cell has multiple sequences assigned to it, it will be counted multiple times
LLAMA_API int llama_get_kv_cache_token_count(const struct llama_context * ctx);
// Returns the number of used KV cells (i.e. have at least one sequence assigned to them)
LLAMA_API int llama_get_kv_cache_used_cells(const struct llama_context * ctx);
// Clear the KV cache
LLAMA_API void llama_kv_cache_clear(

1
prompts/chat-with-qwen.txt Normal file
View file

@ -0,0 +1 @@
You are a helpful assistant.

3
requirements-hf-to-gguf.txt Normal file
View file

@ -0,0 +1,3 @@
-r requirements.txt
torch==2.1.1
transformers==4.35.2

Some files were not shown because too many files have changed in this diff Show more